Proceedings of

WRFER
INTERNATIONAL CONFERENCE
Pune, India

Organized by

Date of Event:
28th June, 2025

In Association with

Corporate Address
IRAJ INTERNATIONAL
Address: IFZA Business Park, DDP Premises Number 49839 001
A1-3641379065, Dubai, UAE
Mail: [email protected], www.iraj.in
Publisher: IRAJ INTERNATIONAL

2025, WRFER International Conference, Pune, India

ISBN: 978-93-90150-25-0
Edn: 511

No part of this book can be reproduced in any form or by any means without prior written
permission of the publisher.

Disclaimer: Authors have ensured sincerely that all the information given in this book is
accurate, true, comprehensive, and correct right from the time it has been brought in writing.
However, the publishers, the editors, and the authors are not to be held responsible for any
kind of omission or error that might appear later on, or for any injury, damage, loss, or
financial concerns that might arise as consequences of using the book.

Type set & Printed by:

ITR Group
Khandagiri, Bhubaneswar
About IRAJ International:

IRAJ International is an advanced global organization for the Researchers & Scholars "To promote the
Progress of Science, Technology and Management" by displaying their knowledge in the vicinity of Science
Technology and Management for the service of mankind and the advancement of the general welfare. We have
been standing tall in the field of Research & Technology since 2013, our objective is to provide a platform
which will promote, encourage and support Scholars, Researchers & Professionals to carry and accomplish their
research work.
IRAJ International connects Global innovators, Scientists, Engineers, Doctors, Management professionals,
industry practitioners by conducting and sponsoring technical meetings, conferences, symposia, and exhibitions
globally. IRAJ International Previously known as Institute of Research and Journals-IRAJ is one of the largest
Professional bodies for Research Scholars worldwide, established in 2013, with headquarters at Dubai, UAE and
branch office at Bhubaneswar, India. IRAJ International reach extends to every part of the globe, with more than
60000 professional members and 50000 student members.

Objective of IRAJ International:

 To provide a world class platform to researchers to share the research findings by organizing
International/National Conferences.
 To use the research output of the conference in the class room for the benefits of the students.
 To encourage researchers to identify significant research issues in identified areas, in the field of Science,
Engineering, Technology and Management.
 To help dissemination of their work through publications in a journal or in the form of conference
proceedings or books.
 To encourage regional and international communication and collaboration; promote professional interaction
and lifelong learning; recognize outstanding contributions of individuals and organizations; encourage
scholar researchers to pursue studies and careers in circuit branches and its applications.
 To set up, establish, maintain and manage centers of excellence for the study of /on related subjects and
discipline and also to run self-Supporting projects for the benefit of needy persons, irrespective of their
caste, creed or religion.

 IRAJ CPD Events: IRAJ International is a globally recognized organization dedicated to fostering
innovation, research, and professional development. One of its key initiatives is organizing continuing
Professional Development (CPD) accreditation events, which serve as a platform for professionals,
academicians, and researchers to enhance their knowledge, skills, and career prospects.

For More Details: www.iraj.in

About WRFER:

WRFER is World's largest Non-profitable professional association registered under India Trust Act(1882) meant
for research development and promotion in the field of engineering and technology. WRFER is a paramount
body which has brought technical revolution and sustainable development of science and technology. The
WRFER-forum constitutes of professional wizards and overseas technical leaders who have left no stones
unturned to reinforce the field of science, engineering and technology. The Institute conduct technical
conferences, seminars and workshop at different parts of country to reduce the gap between curriculum and their
practical implementation among students and research scholars. Today WRFER is one of the leading publisher
of research papers in its high quality peer reviewed journals, proceeding and research magazine. The Institute
provides a brilliant scope of research and development to geniuses and wizards working in the field of
engineering by providing financial aids by which economic constraints won't retaliate the technical growth and
research development.

The institute is supported by its International Advisory Board(IAB) which is not limited to the subcontinent
rather we have geniuses from middle east to boost our organization. Today WRFER is collaborated with leading
technical institutes including Universities and research centers that promotes the exchange of technical
knowledge among our students.

For More Details: www. wrfer.org

 Conference Committee
Program Chair:

Dr. P. Suresh
M.E, Ph.D. Professor and Controller of Examinations,
Karpagam College of Engineering.,
Coimbatore, India.

Mr. Bijan Kumar Barik

Managing Director
IRAJ International, Dubai, UAE

Mr. Deepak Swain,

Conference Convener, WRFER
Mob: +91-8895188931

Mr. Manas Ranjan Prusty

Publication and Distribution Head

INTERNATIONAL ADVISORY MEMBERS

Prof. Goodarz Ahmadi,

Professor, Mechanical and Aeronautical Engineering, Clarkson University, USA

Dr Chi Hieu Le,

Senior Lecturer, University of Greenwich. Kent ME4 4TB. United Kingdom

PROF. (ER.) Anand Nayyar

Department of Computer Applications & I.T.KCL Institute of Management and Technology, Jalandhar
G.T. Road, Jalandhar-144001,Punjab, India.

Prof. R. M. Khaire,
Professor,Dept. Of Elex. and Telecommunication,B, V University,India

Dr.P.Suresh,
Professor,Karpagam College of Engineering,Coimbatore,Tamilnadu

Mark Leeson
Associate Professor (Reader)
Area of Expertise:nanoscale communications,
evolutionary algorithms, network coding and communication systems

Dr. P. K. Agarwal
Professor,Deptt. of Civil Engineering, MANIT Bhopal ,Ph. D: IIT Kanpur
M.E: Civil Engg.IIT Roorkee, Membership: Indian Road Congress (IRC), Institute of Urban Transport (IUT)

Shahriar Shahbazpanahi
Islamic Azad University,
Department of Civil Engineering, Sanandaj, Kurdistan, Iran, PhD (Structural Engineering),
University Putra Malaysia, Malaysia,2009-Present

Harun Bin Sarip

Head of Research and Innovation, Dept, UniKL-MICET
Doctorate: Université de La Rochelle, France
Member : International Society of Pharmaceutical Engineer, Singapore Chapter

Dr. Md. Al-Amin Bhuiyan

Associate Professor
Dept. of Computer Engineering, King Faisal University
Al Ahssa 31982, Saudi Arabia

Prof. (Er.) Anand nayyar

Department of Computer Applications & I.T.
KCL Institute of Management and Technology, Jalandhar
G.T. Road, Jalandhar-144001, Punjab, India

Prof. Aleksandr Cariow

institution or Company: West Pomeranian University of
Technology, Szczecin
Dr. P. K. Agarwal
Professor,Deptt. of Civil Engineering, MANIT Bhopal ,Ph. D: IIT Kanpur
M.E: Civil Engg.IIT Roorkee, Membership: Indian Road Congress (IRC), Institute of Urban Transport (IUT)

Dr. VPS Naidu

Principal Scientist & Assoc. Prof., MSDF Lab, FMCD
CSIR - National Aerospace Laboratories, Bangalore, India

Mr. P. Sita Rama Reddy

Chief Scientist ,Mineral Processing Department, CSIR - Institute of Minerals & Materials Technology
Bhubaneswar,India, M.Tech. (Chem. Engg., IIT, KGP)

Dr.P.C.Srikanth,
Professor & Head, E&C Dept, Malnad College of Engineering,Karnataka
Senior Member IEEE, Secretary IEEE Photonics Society,
M.Tech: IIT, Kanpur, Ph.D: In IISc Photonics lab

Prof. Lalit Kumar Awasthi,

Professor,Department of Computer Science & Engineering
National Institute of Technology(NIT-Hamirpur),
PhD, IIT, Roorkee, M. Tech, IIT, Delhi

Dr. Chandra Mohan V.P.

Assistant Professor, Dept. of Mech. Engg., NIT Warangal,
Warangal. Ph.D : Indian Institute of Technology(IIT),Delhi
M.B.A: Alagappa University

Prof. I.Suneetha,
Associate Professor,Dept. of ECE,AITS,Tirupati,India

Dr.s. Chandra Mohan Reddy,

Assistant Professor (SG) &Head,Dept. of Electronics & Communication Engineering,JNTUA College of
Engineering,Pulivendula,Ph.D,J.N.T. University Anantapur, Anantapuramu

Gurudatt Anil Kulkarni,

I/C HOD E&TC Department,MARATHWADA MITRA MANDAL’S POLYTECHNIC

Pasuluri Bindu Swetha

Dept. Of ECE,Stanley college of Engineering & Technology for Women,Hyderabad, India


 TABLE OF CONTENTS
Sl. No. TITLES AND AUTHORS Page No.

01. AI-Driven Placement Prediction and Insights 1-4

 Dr Manju Pawar, Sumedh Choudhari, Soham Phutane, Anushka Jadhav, Tejas
Bhor

02. Fashion Advisor Model Based on Skin Tone and Body Type 5-8
 Dr. Manju Pawar, Madhura Rajesh Ingole, Piyush Rajendra Bharambe,
Ishwari Ranjit Jagtap

03. “AI Powered Legal Decision Making Assistant System” 9-14

 Dr.Manju Pawar, Sumit Jadhav, Ameya Paigude

04. A Case Study of Synthesis in VLSI Design Using Fusion Compiler 15-19
 Sai Raju Nuguri, Rk Sunkaria

05 Performance Analysis of Double-Gate Staggered Heterojunction TFET Biosensor 20-24

 Rupali, Dr B S Saini

06. Medi Molecule: An AI-Powered Platform for Accelerating Drug Discovery 25-32
Through Molecule Generation and Real-Time Collaboration
 Solapuram Shreeya, Dr. R. M. Noorullah, Annaldas Shivani

07. Multimodal AI in Agriculture: A Fusion of Speech and Image-Based Disease 33-38

Detection
 Dr. Manju Pawar, Tanvi Jadhav, Ankita Satav, Puja Wavhal, Samarth Kargale

08. Advanced IoT Based Rotatable Railway Foot Over Bridge 39-43
 Kamisetti S. L. Sai Greeshmanth, Pampana Siva Harini, Khilari Abhiram,
Dulipudi Esther Rani, Mrs. Chenna V. Kiranmayi

09. Ensemble Hybrid Deep Learning Framework for Fake News Detection Using 44-55
RoBERTa and Explainable AI
 Lakshmi Holla, Dr. Kavitha K.S


 EDITORIAL

It is my proud privilege to welcome you all to the International Conference at Pune,

India. I am happy to see the papers from all part of the world and some of the best paper
published in this proceedings. This proceeding brings out the various Research papers from
diverse areas of Science, Engineering, Technology and Management. This platform is
intended to provide a platform for researchers, educators and professionals to present their
discoveries and innovative practice and to explore future trends and applications in the field
Science and Engineering. However, this conference will also provide a forum for
dissemination of knowledge on both theoretical and applied research on the above said area
with an ultimate aim to bridge the gap between these coherent disciplines of knowledge. Thus
the forum accelerates the trend of development of technology for next generation. Our goal
is to make the Conference proceedings useful and interesting to audiences involved in
research in these areas, as well as to those involved in design, implementation and operation,
to achieve the goal.

I once again give thanks to the IRAJ International, WRFASE & SARC for organizing
this event in Pune, India. I am sure the contributions by the authors shall add value to the
research community. I also thank all the International Advisory members and Reviewers for
making this event a Successful one.

Editor-In-Chief
Dr.P. Suresh
M.E, Ph.D. Professor and Controller of Examinations,
Karpagam College of Engineering,
Coimbatore, India


 AI-DRIVEN PLACEMENT PREDICTION AND INSIGHTS
1
DR MANJU PAWAR, 2SUMEDH CHOUDHARI, 3SOHAM PHUTANE, 4ANUSHKA JADHAV,
5
TEJAS BHOR
1
Department of Artificial Intelligence and Data Science, Zeal College of Engineering and Research, Maharashtra, India
E-mail: [email protected], [email protected], [email protected],
4
[email protected], [email protected]

Abstract - A comprehensive, data-driven platform called "AI-Driven Placement Prediction and Insights" was created to help
academic institutions and students better understand and improve placement preparedness. The system combines a number
of contemporary technologies to provide precise placement forecasts and useful information. Flutter Web was used in the
development of the user interface, making it responsive and fluid.

Primarily, the platform utilizes a Random Forest algorithm-based machine learning model that was trained on a variety of
student characteristics, such as academic performance (10th, 12th, and graduation grades), quiz scores, technical scores (out
of 20), logical reasoning, verbal ability, quantitative aptitude, the number of projects and internships completed, and mastery
of critical skills like SQL, DSA, Java, Python, C++, ML, AI, GenAI, Tableau, ReactJS, NodeJS, and more. Real-time
predictions based on the most recent student inputs are ensured by exposing the model through a Flask API that is
implemented on Render.

The creation of comprehensive Power BI dashboards is made possible by the system's safe management and storage of all
data. These dashboards help administrators and students identify placement trends, skill gaps, and overall performance data
by providing relevant visuals.
To ensure scalability and accessibility, the complete system is integrated into the Flutter Web application, and Firebase
Hosting is used for the final release. This project demonstrates the usefulness of AI and ML in the educational field as well
as the ability of full-stack integration to provide insightful, engaging, and significant solutions.

Keywords - Placement Analysis, Placement Prediction, Machine Learning, Data Visualization, Data-Driven

I. INTRODUCTION recommendations for improvement. Additionally, the

tool ensures scalability, data security, and ease of
In today’s competitive job market, student access, making it suitable for diverse educational
placements play a critical role in shaping the future of institutions.
graduates and determining the reputation of
educational institutions. However, predicting This project bridges the gap between academic
placement outcomes and identifying the factors preparation and industry expectations, enabling
influencing employability remain challenging tasks institutions to enhance their placement strategies and
due to the dynamic nature of industry demands and helping students unlock their career potential.
student profiles.
II. HELPFUL LITERATURE REVIEW
This project, Placement Analysis and Prediction
Using Machine Learning, addresses these challenges The development of predictive systems for student
by leveraging the power of data science and placements has seen significant advancements in
predictive modeling. recent years, primarily due to the application of
machine learning (ML) techniques. Researchers like
By analyzing historical placement data and current Milind Ruparel et al. (2024) and Advait Joshi et al.
student attributes, the system provides actionable (2024) have demonstrated the effectiveness of
insights into placement trends and predicts the supervised learning algorithms such as Logistic
likelihood of students securing job offers. The project Regression, Random Forest, and Decision Trees in
aims to empower students and academic predicting placement probabilities. These algorithms
administrators to make informed decisions by leverage features like academic performance,
detecting talent shortages and aligning their efforts technical skills, and extracurricular activities to
with industry requirements. enhance prediction accuracy.

The system incorporates machine learning algorithms Several studies emphasize the importance of feature
to analyze features such as programming skills, selection and data preprocessing in improving model
academic performance, extracurricular involvement, performance. For instance, Rushikesh Joshi et al.
and recruitment patterns. Through visualizations and (2024) highlighted how including factors like
predictive analytics, the platform offers a user- internships and project work can significantly refine
friendly interface that highlights trends and provides predictions. Similarly, Vinod Agrawal et al. (2023)

Proceedings of WRFER International Conference, Pune, India, 28th June, 2025

1
 AI-Driven Placement Prediction and Insights

reviewed various ML algorithms, finding Decision Networks, Random Forests, Decision Trees, and
Trees and Random Forests particularly effective in Logistic Regression.
placement prediction.
 Model Training:
Moreover, modern predictive systems aim to go  Splitting the data into training (80%) and testing
beyond mere prediction by offering actionable (20%) sets.
insights. Priyanka Shahane (2022) and Varsha  Using cross-validation to ensure robustness and
Harihar (2020) underscored the importance of avoid overfitting.
integrating visualization tools and analytical features
to identify skill gaps and recommend interventions. 4. Model Evaluation
These systems are not limited to predictions but also  Metrics: Evaluating the models using metrics
provide strategic guidance to institutions for like:
enhancing placement outcomes.  Accuracy
 Precision
The literature collectively underscores the potential of
 Recall
ML-driven predictive tools in transforming traditional
 F1-score
placement processes into data-driven frameworks.
 Comparison: Comparing the performance of
This lays a strong foundation for the proposed
various models to select the best-performing one
"Placement Analysis and Prediction" project, which
for deployment.
aims to implement a robust and accurate system
tailored for real-world educational applications.
5. Insights and Recommendations:
III. METHODOLOGY
The implementation of the Placement Prediction and Branch performance insights
Analysis project can be divided into several key  Evaluates effectiveness of individual branches.
phases, each encompassing specific tasks and  Identifies top performers and areas for
objectives. These phases guide the systematic improvement.
development and deployment of the project:
Yearly placement analysis
1. Data Collection  Examines placement trends and statistics across
 Sources: Historical placement data from the different years.
college's database, including:  Reveals long-term patterns and growth in
 Academic scores (e.g., GPA). placements.
 Technical skills and certifications.
 Internship and project details. Company insights: leading placement volume
 Identifies companies recruiting the most students
 Extracurricular activities.
(placement volume).
 Placement outcomes.
 Format: Data is collected in a structured format
Uncovering trends by location, department, And
(e.g., CSV files).
role:
 Analyzes placement patterns based on
2. Data Preprocessing
geographic location, academic department, and
 Data Cleaning: specific job roles.
 Handling missing values (e.g., using mean/mode
 Helps understand skill demand and curriculum
imputation). needs.
 Removing or correcting outliers.
6. Visualization and Reporting
 Data Transformation: Visualization Tools: Creating dashboards to display
Encoding categorical data (e.g., programming placement probabilities, influencing factors, and
languages known) into numerical form. trends using tools like PowerBI or Python
visualization libraries.
 Feature Engineering:
Recognizing and developing novel traits that might Reports: Generating reports summarizing placement
affect placement results. insights and predictions.

3. Model Development 7. Deployment

 Platform: Deploying the system as a web or
 Algorithm Selection: mobile application for ease of access.
The selection of algorithms for classification  Tools: Using platforms like Flask, Flutter for
problems is dependent on factors including Neural web deployment, with cloud support (Render).

Proceedings of WRFER International Conference, Pune, India, 28th June, 2025

2
 AI-Driven Placement Prediction and Insights

8. Testing and Validation

 User Testing: Collecting feedback from students
and administrators to refine the system.
 Performance Validation: Ensuring the model
performs well on unseen data and aligns with
real-world requirements.

IV. RESULTS

Random Forest achieved a perfect accuracy of 100%,

with precision, recall, and F1-score all equal to 1.00
for both classes. While this may indicate strong
predictive capability, such perfect results could also
suggest possible overfitting, especially in real-world
applications.

Proceedings of WRFER International Conference, Pune, India, 28th June, 2025

3
 AI-Driven Placement Prediction and Insights

coherent platform, this method is both practical and

expandable. By closing the gap between academic
success and industry readiness, it creates the
foundation for more astute, data-driven institutional
planning and career advising.

FUTURE WORK

While the current model shows promising results,

several avenues exist for further improvement:
1. Addressing Class Imbalance: Future iterations
should implement advanced techniques such as
SMOTE, ADASYN, or cost-sensitive learning to
mitigate class imbalance and improve
performance on minority classes.
2. Incorporating Real-Time Data: Integrating
dynamic or real-time student performance data
(e.g., ongoing project performance, recent
certifications) could enhance prediction
accuracy.
3. Model Interpretability: Introducing explainable
AI (XAI) methods can help users understand the
rationale behind predictions, increasing
transparency and trust in the system.
4. Deployment at Scale: Future development can
focus on making the system scalable across
multiple institutions, integrating APIs, cloud-
V. CONCLUSION AND FUTUREWORK based storage, and robust security mechanisms.
5. Personalized Recommendations: Enhancing the
The dataset's class imbalance issue is demonstrated system to provide personalized roadmaps for
by the consistently declining macro F1-scores for all individual students based on their unique
model but Random Forest. Overall, the study shows strengths and weaknesses can significantly boost
that Random Forest is the most robust and complete their placement chances.
model for this prediction task. The Random Forest
classifier outperformed the other machine learning REFERENCES
techniques with perfect results in accuracy, precision,
recall, and F1-score. This demonstrates its high [1] Joshi, A. Hutke, A. Thete, D. Aware, S. Rawale, and N. A.
Kandalkar, "College placement prediction and analysis using data
predictive capacity, but it also raises concerns about science," International Research Journal of Modernization in
potential overfitting, which requires further Engineering Technology and Science, vol. 6, no. 3, pp. 4336–
investigation. The relatively good performance of 4340, Mar. 2024. [Online]. Available:
www.irjmets.com:contentReference[oaicite:0]{index=0}.
other models, such as Logistic Regression, XGBoost, [2] M. Ruparel and P. Swaminarayan, "Student placement prediction
and SVM, but their difficulties with class imbalance using various machine learning techniques," International Journal
brought attention to the importance of dataset of Intelligent Systems and Applications in Engineering, vol. 12,
no. 3, pp. 2107–2113, Mar. 2024. [Online]. Available:
balancing and model tuning. www.ijisae.org:contentReference[oaicite:1]{index=1}.
The study also highlights how important it is to assess [3] K. Srinivas, N. S. Yadav, P. A. Somashekar, and S. K. R.,
and display model results using intuitive dashboards "Students placement prediction using machine learning,"
International Journal for Research in Applied Science and
and analytics. In addition to helping administrators Engineering Technology (IJRASET), vol. 8, no. 5, pp. 2771–
make data-driven strategic choices, these visual aids 2774, May 2020. [Online]. Available:
assist students in identifying their areas of strength www.ijraset.com:contentReference[oaicite:2]{index=2}.
[4] V. S. Agrawal and S. Shelke Kadam, "Review on predictive
and progress. Project experience, internships, analysis of placement of students using machine learning
programming skills, and academic accomplishment algorithms," in ICAMIDA 2022, Aurangabad, MH, India: MGM
were shown to be significant indications that closely University, 2023, pp. 23–30. doi: 10.2991/978-94-6463-136-
4_4(3).
matched the demands of the real industry. [5] P. Shahane, "Campus Placements Prediction & Analysis using
All things considered, this study demonstrates how Machine Learning," in 2022 International Conference on
educational institutions may apply these findings to Emerging Smart Computing and Informatics (ESCI), Pune, India,
Mar. 2022, pp. 1-5, doi: 10.1109/ESCI53509.2022.9758214(1).
improve placement outcomes and validates that [6] V. K. Harihar and D. G. Bhalke, "Student Placement Prediction
machine learning is a valid technique for predicting System using Machine Learning," SAMRIDDHI: A Journal of
placement prevalence. Because prediction, analysis, Physical Sciences, Engineering and Technology, vol. 12, no. SI-2,
pp. 85-91, 2020, doi: 10.18090/samriddhi.v12iS2.17(7).
and recommendation are all integrated into a single,


Proceedings of WRFER International Conference, Pune, India, 28th June, 2025

4
 FASHION ADVISOR MODEL BASED ON SKIN TONE AND BODY TYPE
1
DR. MANJU PAWAR, 2MADHURA RAJESH INGOLE, 3PIYUSH RAJENDRA BHARAMBE,
4
ISHWARI RANJIT JAGTAP

Department of Artificial Intelligence and Data Science, Zeal College of Engineering and Research, Maharashtra, India
E-mail: [email protected], [email protected], [email protected],
4
[email protected]

Abstract - In an era of personalized consumer experiences, the need for tailored fashion advice has grown significantly.
This study offers a Fashion Advisor Model created especially for individual customers at a designer studio, offering
suggestions based on a mix of body type, skin tone, and cultural fashion tastes. Using machine learning techniques like
GANs, CNNs, and decision trees, together with skin color recognition via the HSV color model, the model provides
real-time individualized clothing recommendations. It customizes its suggestions based on user profiles using picture
recognition and personal input, offering outfit options for both corporate and informal settings. These suggestions are
further improved over time by a feedback process, which further increases user happiness and accuracy. This model
is a prime example of how sophisticated AI can enhance fashion advice and provide consumers a customized,
interactive, and one-of-a-kind style experience.

Keywords - Fashion Recommendation, Skin Tone Detection, Body Type Classification, GANs, Personalized Styling,
CNN, HSV Color Mode

I. INTRODUCTION and random forests, which categorize clothing

according to user similarities, further enable
Personalized fashion recommendations are personalized recommendations. A mix of human
revolutionizing how customers approach their style, input and image recognition is used to gather user
especially in designer studios, where individuality is data, providing the model with current and accurate
paramount. Designed to be utilized by individual information about each customer’s profile. When
clients in such circumstances, the pro- posed Fashion the model is finished processing, it displays
Advisor Model aims to provide a smart styling personalized clothing suggestions in the form of
assistant that respects both the functional and created pictures. With an easy-to-use interface
aesthetic elements of fashion. The model considers powered by HTML, CSS, and JavaScript, and Flask
users’ unique skin tones, body types, and cultural for backend processing, the system offers a seamless
preferences in an effort to offer recommendations that user experience.
not only match with cultural context but also
highlight the wearer’s natural characteristics. The A feedback mechanism also enables users to
model classifies skin tone using the Hue, Saturation, respond to and edit their suggestions, promoting
Value (HSV) color model, which allows for a ongoing enhancements to the model’s precision and
sophisticated way to discover undertones that might applicability. This Fashion Advisor Model
influence color harmony in clothing choices. demonstrates the possibility of AI-driven fashion
guidance in high-end, personalized retail
The classification of body types follows common environments. By combining technological
archetypes, such as Rectangle, Inverted Triangle, proficiency with fashion understanding, it provides
Hourglass, Apple, Pear, Spoon, and Diamond, to customers with personalized, culturally appropriate,
ensure that the costumes produced fit a variety of and contextually appropriate style solutions.
shapes and sizes. Business dress, casual styles, and
culturally acceptable clothing are all discussed to II. RELATED WORK
ensure that the recommendations are suitable for a
range of situations. Using a variety of methods, Personalized fashion recommendation systems have
including Generative Adversarial Networks (GANs), garnered a lot of interest as current research focuses
the Fashion Advisor Model technically visualizes on enhancing user experience through tailored
outfits and creates style combinations that satisfy recommendations based on individual preferences,
client specifications. Content-based and collaborative skin tone, and body form. Previous studies have
filtering techniques are commonly used by demonstrated that a variety of methods, including
recommendation systems to analyze user preferences content-based filtering, collaborative filtering, and
and enhance recommendations. To increase the deep learning approaches, may accomplish
accuracy of image recognition for users’ skin tone customization. Despite these advancements,
and body type, Convolutional Neural Networks challenges such as the cold-start problem, data
(CNNs) are integrated into the model. Algorithms imbalance, and accurately determining body shapes
such as k-Nearest Neighbors (KNN), decision trees, and skin tones persist

Proceedings of WRFER International Conference, Pune, India, 28th June, 2025

5
 Fashion Advisor Model Based on Skin Tone and Body Type

1. COLLABORATIVE AND CONTENT- body type classification, it seeks to overcome

BASED FILTERING challenges related to data requirements and
personalization accuracy. The integration of GANs,
Collaborative filtering and content-based filtering are Decision Trees, and Random Forests further enhances
commonly employed in fashion recommendation the model’s adaptability, aiming to deliver a more
systems. Kun- cham et al. (2024) developed a inclusive and user- centric recommendation system.
comprehensive fashion recommender system that
combines these methods, addressing issues of user III. LITERATURE REVIEW
satisfaction and preference prediction. Vinitha et al.
(2024) incorporated collaborative filtering and real- Personalized fashion recommendation systems have
time behavior analysis in a system that adapts to user seen significant advancements, focusing on tailoring
preferences. However, both systems face difficulties suggestions to user preferences, body types, and skin
with new users and require extensive datasets to tones. Kuncham et al. (2024) and Vinitha et al.
achieve accuracy. (2024) employed collaborative and content-based
filtering, highlighting their efficiency in generating
2. SKIN TONE DETECTION AND COLOR recommendations. However, these systems face
ANALYSIS limitations such as the cold-start problem and the
need for extensive datasets to achieve accuracy. Body
Skin tone detection is integral to creating inclusive type classification has been a focal point in several
fashion recommendations, as it allows for color studies. Rode et al. (2024) utilized machine learning
palettes that enhance an individual’s natural algorithms to analyze body shapes, while Wazarkar et
complexion. Studies by Chakraborty et al. (2021) and al. (2022) leveraged deep learning for similar
Daithankar et al. (2014) examined various color purposes. Despite these efforts, challenges in
spaces, such as RGB and HSV, to optimize skin tone obtaining accurate body type data and dependence on
detection in recommendation models. Their findings substantial training datasets remain persistent
underscore the efficacy of CNN-based models for skin obstacles. Skin tone detection has also been explored
detection, but they also highlight limitations in to enhance inclusivity in fashion recommendations.
handling diverse lighting conditions and varying skin Chakraborty et al. (2021) and Daithankar et al. (2014)
tones. examined the efficacy of color models such as RGB
and HSV for skin tone analysis. While CNN-based
3. DEEP LEARNING IN FASHION approaches improved accuracy, they struggled with
RECOMMENDATIONS lighting variations and the diverse range of skin
tones. Deep learning methods have enriched fashion
Deep learning models have been effective in recommendation systems by processing visual
processing large amounts of image and text data to features and text data. For instance, Sivaranjani et
improve fashion recommendation accuracy. al. (2023) employed CNNs, KNN, and RESNET50 to
Sivaranjani et al. (2023) used CNNs, KNN, and recommend items based on image analysis, while
RESNET50 to classify and recommend fashion items Yin et al. (2023) explored Auto ML techniques to
based on visual features. However, these models are enhance model components. These methods face
often limited by data imbalance issues and the scalability challenges and issues with imbalanced
inability to account for non-visual preferences. The datasets, limiting their broader applicability. Hybrid
study by Wazarkar et al. (2022) also illustrated how approaches integrating advanced algorithms like
deep learning models can categorize body types, but GANs and filtering methods have shown promise in
their dependency on vast datasets poses scalability creating user-centric recommendation systems. Tu
challenges. and Dong (2010) pioneered early models combining
multimedia mining with color analysis, while Iso and
4. AUTO ML AND HYBRID MODELS: Shimizu (2021) incorporated personal style analysis
for more nuanced recommendations. However,
More recent research, such as Yin et al. (2023), adapting to rapidly evolving fashion trends remains a
explores Auto ML techniques to enhance deep significant hurdle. The proposed Fashion Advisor
recommender systems by optimizing model Model builds upon these foundations by addressing
components and balancing exploration with limitations in existing systems. By integrating
exploitation. However, while Auto ML shows advanced algorithms such as GANs, Decision Trees,
promise, it struggles to streamline system and CNNs, and considering cultural preferences, the
components effectively and adapt to rapidly changing model aspires to deliver highly personalized and
trends without continuous model refinement. The inclusive recommendations. Its focus on real-time
proposed Fashion Advisor Model builds upon these adaptability and user feedback mechanisms marks a
studies by addressing the limitations identified in critical step toward improving accuracy and
existing systems. By incorporating collaborative and satisfaction in fashion advisement systems.
content-based filtering with advanced skin tone and

Proceedings of WRFER International Conference, Pune, India, 28th June, 2025

6
 Fashion Advisor Model Based on Skin Tone and Body Type

IV. METHODOLOGY tone classification through HSV will be executed

directly within this interface, enhancing processing
1. Data Collection and Input Mechanism speed and user experience.
Data will be gathered through both manual entry and
image recognition methods. Users will select their V. RESULTS AND ANALYSIS
body type from predetermined options, such as
Rectangle, Inverted Triangle, Hourglass, Apple, Pear, 1. Model Evaluation
Spoon, and Diamond. Image recognition will use To validate model performance, the Fashion Advisor
CNNs to enhance body type detection accuracy and Model will be tested on an image dataset covering
incorporate user-uploaded photos. The HSV color various body types and skin tones. Key metrics such
model will support a detailed classification of skin as accuracy in body type and skin tone detection,
tone to guide recommendations for the most as well as user satisfaction based on feedback, will
complementary colors. be used to evaluate the effectiveness of each
algorithm. GAN-generated outfit combinations will
2. Algorithms and Model Architecture be evaluated based on visual quality and user
The recommendation model will leverage a acceptance rates.
combination of machine learning techniques and
deep learning algorithms: 2. Comparative Analysis
 GANs (Generative Adversarial Networks): The model will be compared to existing
Employed for generating diverse outfit recommendation systems lacking skin tone and body
combinations based on body type and skin tone, type personalization. Traditional methods that focus
GANs will enable the creation of visually on user purchase history alone are expected to
appealing outfit options. underperform in terms of relevance and user
 Content-Based and Collaborative Filtering: satisfaction, especially within culturally rich
Content-based filtering will analyze clothing demographics. This comparison will highlight the
features that match user attributes, while effectiveness of combining GANs with machine
collaborative filtering will consider peer learning algorithms for enhanced personalization.
preferences, enhancing recommendations
through combined insights. 3. Discussion
 K-Nearest Neighbors (KNN): Used for locating The Fashion Advisor Model demonstrates potential
fashion items with similar attributes, KNN will for increasing user satisfaction by offering
further support accurate and efficient recommendations that respect and celebrate
recommendations. individual body types and skin tones. Integrating
 Convolutional Neural Networks (CNNs): CNNs cultural attire considerations adds a distinct ad-
will be critical for both skin tone classification vantage in markets like India, where fashion
and body type verification, using uploaded preferences are strongly influenced by traditional
images to better detect and suggest suitable attire and evolving modern styles. By considering
styles. regional attire, the model aligns with cultural
 Decision Trees and Random Forests: To increase sensitivities and enhances users’ confidence in their
the model’s robustness, Decision Trees and clothing choices. This approach challenges existing
Random Forests will aid in determining the best fashion recommendation systems, which often miss
classification patterns for different body types, important details by relying solely on general
based on user feedback and model training. attributes. The GAN- generated combinations and
diverse filtering techniques elevate the model’s
3. User Feedback Loop capacity for personalization, showcasing AI’s ability
User feedback will be collected through images and to adapt to highly individualized fashion requirements
detailed comments, allowing the model to refine
recommendations over time. By continuously 4. Conclusion and Future Work
learning from user interactions, the model adapts to The proposed Fashion Advisor Model addresses a
evolving preferences and changing trends. critical gap in the current fashion recommendation
Feedback will also help address the cold-start landscape by providing culturally relevant, body-type,
problem by allowing the model to learn from each and skin-tone specific recommendations. Based on AI
unique interaction. and machine intelligence, the idea facilitates
decision-making for customers in designer studios
4. Web-Based Interface Development while promoting individual expression. Future
A Flask-based interface supported by HTML, CSS, improvements will include real-time trend monitoring
and JavaScript will enable users to input their data and and expand the model to include more body types and
interact with recommendations intuitively. Users will outfit combinations to ensure that suggestions align
be able to manually select their body type, upload with current fashion trends. By enhancing
images, and view generated outfit combinations. Skin personalization, this approach encourages inclusivity

Proceedings of WRFER International Conference, Pune, India, 28th June, 2025

7
 Fashion Advisor Model Based on Skin Tone and Body Type

and self- assurance in users, making purchasing Nagell, “Fashion Recommendation System” , IJRASET,
2022
more enjoyable and personalized. The model’s
[3] M Vinitha, Dr.B. Nagarajanaik , Mallikarjuna Nandi , C Naga
success in the Indian market might set a benchmark Sri Charan, K Priyanka, “Fashion Recommendation System”,
for fashion advising systems elsewhere, particularly IRJAEH, 2024
in regions with similar degrees of cultural diversity [4] Tangseng P., Okatani T. , “Toward explainable fashion
recommen- dation”,IEEE, 2020
and a need for tailored guidance
[5] Pujahari A. Sisodia, D.S. , “Model-based collaborative
filtering for recommender systems: An empirical survey”,
REFERENCES IEEE, 2020
[6] Sagar D. Garg, J.Kansal, P.Bhalla, S. Shah, R.R. Yu,
[1] Samit Chakraborty, Md. Saiful Hoque, Naimur Rahman Jeem, “Personalized Outfit Recommendation Scheme with
Manik Chandra Biswas, Deepayan Bardhan and Edgar Attribute-wise Interpretability”, IEEE , 2020.
Lobaton, “ Fashion Recommendation Systems” , Models and [7] Pujahari A. Sisodia, D.S., “Model-based collaborative
Methods: A Review, Research Gate,2024 filtering for recommender systems”, ICPC2T, 2020
[2] Aneesh K, P. V. Rohith Kumar, Sai Uday Nagula, Archana



Proceedings of WRFER International Conference, Pune, India, 28th June, 2025

8
 “AI POWERED LEGAL DECISION MAKING ASSISTANT SYSTEM”
1
DR.MANJU PAWAR, 2SUMIT JADHAV, 3AMEYA PAIGUDE
1,2,3
Department of Artificial Intelligence & Data Science, Zeal College of Engineering and Research, Narhe
Email: [email protected], [email protected], [email protected]

Abstract - The integration of Artificial Intelligence(AI) in the field of law is transforming conventional practices, providing
greater efficiency, accuracy and accessibility. The paper illustrates the design and development of an AI-based Legal
Decision-Making Assistant System. It is made to support legal professionals , judges and policymakers in making well-
researched and consistent decisions. The system uses Natural Language Processing(NLP), Machine Learning (ML) and legal
knowledge graphs to process massive amounts of legal texts, precedents , legislation and case law. Through human-like
reasoning and understanding simulated by the system, the assistant can deliver case-specific results, predict possible
,outcomes , suggest relevant legal rules and detect meaningful patterns. The solution not only speeds up legal research but
also offers transparency and fairness in judicial decision -making .Ethical safeguards and explainable AI(XAI) principles are
incorporated in the system to ensure accountability and trust. This research illustrates how AI can revolutionize legal
reasoning and provides directions for future research to incorporate it into judicial and legal advisory systems.

I. INTRODUCTION framework for an AI – based legal assistant and

evaluate its effectiveness in actual legal applications.
The judicial process plays critical role in the Through the incorporation of AI in the legal field, this
satisfaction of justice, quality in society. Nevertheless research aims to contribute to the development of a
increasing , dependency in cases , legal complexity in more effective, data-driven and accessible legal
proceeding , labour intensive manual research have system.
resulted in efficiencies the judicial decision making
process. II. BACKGROUND & RELATED WORK

According to the most recent reports courts in The legal profession is going through a revolutionary
different jurisdictions are seriously backed up, change with the emergence of Artificial Intelligence
stretching case hearing delays and exerting (AI) technologies.Traditionally ,legal decision-
tremendous pressure on legal research take making has relied upon human interpretation of
considerable time, labour , resulting in disparities and statutes,judicial precedents ,and facts of cases an
human fallibility. With the emergence of Artificial exercise time -consuming in nature based on
Intelligence (AI) and Machine Learning (ML) legal expertise and critical analysis.
profession is also changing . AI – based Legal
Decision Making-Assistant System are being The exponential growth of legal data,complexity of
developed to assist lawyers, judges and legal experts cases and requirements of uniform and timely
by automating legal research , case analysis, and decisions ,however ,has made intelligent systems with
judgement prediction. the ability to support decision making processes a
necessity.AI-based legal decision -making support
These systems employ Natural Language Processing sytems are a rapidly developing field ,driven by the
(NLP),DeepLearning and knowledge Graphs to promise of significantly enhancing efficiency and
process vast legal data, find applicable Case laws and access to justice. Legal professionals have to deal
provide data-driven insights to improve decision with a huge amount of information , such as case law,
making. This study presents the development and statues, regulations and contracts.AI can process and
design of an AI- based legal assistant system that analyze such information more effectively . NLP
improves legal efficiency through case analysis plays a key role in analyzing legal texts, extracting
automation, classify of similar legal case and valuable information and comphrehending legal
predictions of probable outcomes based on historical concepts.ML algorithms can be employed to forecast
data. The study also explores the potential of AI in legal outcomes , legal data trends and legal risks.
minimizing human bias, making the law more Legal chatbots can assist in providing simple legal
accessible and generally enhancing the effectiveness information and responding to frequent questions.
of the judiciary. While ,AI provides promising
,solutions ,ethical concerns such as bias in AI models This can enhance access to legal aid and reduce the
and accountability in legal decision making are issues workload of legal professionals. Employing Bert
that require serious attention. models in legal decision making assistant system to
load pre-trained Bert based Question answering
The goal of this research is to critically assess current model. An improved user interface would be needed
AI, models used in legal analytics ,purpose a for practical use.

Proceedings of WRFER International Conference, Pune, India, 28th June, 2025

9
 “AI Powered Legal Decision Making Assistant System”

III. PROPOSED SYSTEM prototyping ,with the loss function managed

internally by the Trainer using cross -entropy for
Here, we will demonstrate a tool that is designed to start and end position classification.
assist lawyers, judges, and researchers in analyzing 5. Evaluation and Metrics: In these project we are
legal cases, summarizing key information , and using evaluation metrics SQUAD v1.1 metric. F1
providing insights based on facts in bid to assist in score also used to measures or overlap between
making informed decisions. It is a natural language prediction and ground truth.
chatbot that communicates with users via a Question- 6. Model Saving and Deployment: Save fine -tuned
Answer(QA) interface. The chatbot has been trained model and tokenizer locally using Save
using a well-annotated legal dataset(for instance, Pretrained ().Integrates with the Gradio for
constitutional law). creating the simple and creative UI for chatbot.

Get the correct legal rules, court cases, and IV. APPROACHES / METHODOLOGY
explanations. Offer initial legal opinions or advice
based on available knowledge .Have human -like A. Legal Data Collection
conversation with a natural language interface. Ensures the constitution ,statues, case law ,legal
commentaries, court judgements sources. Ensures the
information given at a Question Answer Format.

B. Preprocessing and Annotation

Clean unwanted characters, format case names,
footnotes removal .
Breaks large text into semantically significant
chunks(clauses ,sections).

C. Natural Language Processing(NLP)

Grasps the relevant text and retains tokens and
identifies names entities, such as people organizations
,locations , dates andothers , in unstructured text.
Fig 1. Proposed Method Framework
Comprehends clause structure and context of legal
Figure 1 shows the proposed framework for this sentences . Translates user queries and legal content
study. A brief introduction about system components into high dimensional vectors.
are as follows:-
1. Data Collection and Preparation : Legal texts D. Legal Domain Adaptation
(Indian Constitutional law dataset) are gathered Tune the base models (e.g., BERT,T5) with domain -
and cleaned. The data is converted to a specific training datasets such as Indian legal Corpus,
structured form appropriate for indexing and Constitutional QA pairs.
retrieval .
2. Data preprocessing :Tokenize the question and Enhance the context interpretation of legal
context using Bert Tokenizer Fast. Offset terminology.
Mapping align character based answer start with
token based offsets. Dataset Splitting 80:20 ratio E. Answer Generation
for training and evaluation sets. Answer generation is the last and most important step
3. Model Selection : Using the pre-trained models in a legal QA system. Once retrieving the most
like Bert-based-uncased models. Better suitable legal content, the system must generate a
performance with less labelled data. Predicts start consistent ,contextually correct, and legally valid
and end positions of the answer span within the response that users will rely on .Utilizes pre-defined
context. Can be fine tuned to understand domain patterns to reply to predictable questions(articles).
specific terms and structure. Fine tuning enables
the model to adapt to legal domain specific Utilizes a language model(e.g.,GPT-4,T5,BART) to
language ,making it robust automated legal conditionally generate answers on both the question
decision support. of the user and retrieval legal context.
4. Fine Tuning the Model: Fine-tuning was
conducted using the Hugging Face Trainer API, V. RESULTS & ANALYSIS
which provided an abstraction layer over the
training loop, optimizer and scheduler. 1) The performance of the AI based legal assistant
Trainingwas performed with a batch size of 2 to system was measured both in terms of
accommodate ,while gradient accumulations quantitative measures(e.g., reviews from legal
steps were employed to stimulate a larger experts, customers satisfications
effective batch size. A single epoch was used for
Proceedings of WRFER International Conference, Pune, India, 28th June, 2025
10
 “AI Powered Legal Decision Making Assistant System”

Feature Description Result

Legal QA Handling Ability to answer direct legal 65%

questions.

Evaluation Metrics SQUAD Style QA 65%

Exact Match, 58%
F1 Score
2) Qualitative Evaluation Metrics ,SQUAD style 4) The chatbot successfully tracked the context and
QA and Exact Match. provided an appropriate , citied answer.
3) The system correctly identified the article text , 5) Below the graph shows a comparison between
and explained that Article 21 includes the right to Right Case law and Procedural Law +
life, dignity ,privacy ,and legal aid. Definitions.

Fig 2. Accuracy Metrics Graph.

Bert-base-uncased model. We fine tuned this pre-
TECHNIQUES trained model on a domain specific, SQUAD style
legal QA dataset consisting of Indian constitutional
1) Natural Language Processing(NLP) texts and their corresponding question answer pairs.
Fine tuning enables the model to adapt its parameters
Legal Text Parsing & Preprocessing to learn domain specific answer span detection in
legal documents, while preserving its general
Apply tokenization ,sentencesegmentation , named understanding of syntax, semantics.
entity recognition (NER), and part- of -speech
tagging specific to legal texts. 3) Supervised learning with fine-tuning.
Supervised learning with fine-tuning is a specialized
Legal Document Understanding form of machine learning where a pre-trained model
is adapted to a new task using a labelled dataset. The
Utilize models such as Legal Bert, Case Law Bert ,or supervised fine tuning approach significantly reduces
fine-tuned LLMs to analyze contracts, statutes , and the data and computer resources required for high
case law. Provide brief summaries of voluminous accuracy. The use of Hugging Face’s Tainer API
legal documents or judgements. Apply retrieval - allows efficient batching ,gradient accumulation and
augmented QA systems for legal statues and case- epoch based training with real time evaluation using
based queries. SQUAD metrics.

2) Transfer Learning 4) Model Evaluation

Transfer learning is a foundation of deep learning Model evaluation is important phase in developing
technique that can use pre-trained models for decision NLP systems to ensure the model performs
making assistant it can plays major role in pre trained accurately, reliable and generalizes well to unseen
Proceedings of WRFER International Conference, Pune, India, 28th June, 2025
11
 “AI Powered Legal Decision Making Assistant System”

data. In our AI powered legal decision making ,while the F1 score captures partial overlaps by
assistant system, we evaluate the fine tuned Bert computing the harmonic mean of precision and recall
based question answering model using standard over predicted and true answer spans.
extractive QA metrics: Exact Match(EM) and F1
score, based on the SQUAD evaluation framework. 5) Model Serving
The EM metric measures the percentage of Using the frontend for chatbot of Gradio. Creating
predictions that exactly match the truth answers simple attractive UI interface .

LITERATURE SURVEY

Proceedings of WRFER International Conference, Pune, India, 28th June, 2025

12
 “AI Powered Legal Decision Making Assistant System”

CHALLENGES legal databases, and assisting in sound legal

reasoning.
Making a successful AI Powered legal Decision
making Assistant System implies a number of major But this transition doesn’t take place without
challenges: significant challenges. The initiative must navigate
very carefully through the prisms of legal language
1) Complexity and Ambiguity in legal ,dataquality ,differences in jurisdiction ,and the
Languauge. ethical mandate of AI deployment in high -risk
Constitutional language is vague, wordy , and fraught environments. Trust ,transparency ,and human agency
with subtle legal . in decision-making are just as important.
NLP models can error in interpreting legal concepts
or they can oversimplify them. Sample :Words and Multidisciplinary collaboration is necessary when
phrases like “reasonable restrictions” or “due assembling a team of experts with law, artificial
process” can be very context dependent as well as intelligence ethics ,and software engineering
their interpretation . backgrounds in order to make the project a success.
2) Data Quality & Availability of Data. Ongoing verification of user input , and compliance
Data in the form of laws tend to be diverse (cases, with regulations should be the guiding force for the
statues, agreements), unstructured, and country - development process.
specific. In conclusion ,while path to building an AI legal
It challenging to access current and complete assistant is challenging ,it is a promising and
datasets(case law, regulations) because of licensing or worthwhile endeavour. If implemented responsibly ,
paywalls. such systems can significantly contribute to more
3) Interpretability of AI Models accessible and data driven systems.
Deep learning models (particularly LLMs) can give
correct outputs , but it is challenging to explain how ACKNOWLEDGMENT
decisions are reached.
4) Contextual Reasoning We are immensely thankful to every individual who
Legal judgements tend to rely on knowledge of in one way or another contributed to the success of
precedents , jurisdictional principles ,intent , and this project. To begin with, I thank deeply my project
factual distinctions. advisor Prof. Gajanan Bhusare for his incredibly
5) Integration with Existing Legal Workflows useful guidance, constructive comments, and
Law firms or courts might have legacy systems that unending support during the course of the research.
are difficult to integrate with contemporary AI
architectures(PDFs, proprietary tools, etc.) Thanks also to Zeal Institute for making available the
required facilities and infrastructure needed for the
CONCLUSION completion of this research.

In this project we propose that an AI -Based legal Our gratitude goes to our colleagues and peers for
decision making assistant system has a vast potential their contributions to our suggestions and
to transform the legal system by making it more encouragement that was vital in the fine-tuning of the
efficient, accessible to justice , and enhancing project to completion and for achieving the objectives
decision making for legal professionals and of the project in a more holistic way.
institutions. With powerful AI models ,such systems At last, I express my deepest gratitude to my family
are capable of doing routine tasks, handling large and friends for their assistance and motivation

Proceedings of WRFER International Conference, Pune, India, 28th June, 2025

13
 “AI Powered Legal Decision Making Assistant System”

throughout the duration of this project. Their words [3] [P. Vimala Imogen¹, J. Sreenidhi², V. Nivedha][2024] AI-
of strength and patience were very helpful in staying Powered Legal Documentation Assistant.Journal of Artificial
Intelligence and Capsule Networks (ISSN: 2582-
dedicated and within the limits of our objectives. 2012)www.irojournals.com/aicn/DOI:
https://doi.org/10.36548/jaicn.2024.2.007Received:
REFERENCES 08.04.2024, received in revised form: 04.05.2024, accepted:
21.05.2024, published: 01.06.2024
[1] D. A. Pashentsev1, Yu. G.Babaeva2.][2024] Artificial [4] [Sukwoong Choi, Hyo Kang Namil,
intelligence in law-making and lawPashentsev, Dmitrii A., KimgJunsikKim][2023]How Does Artificial Intelligence
Yuliya G. Babaeva. 2024. “Artificial intelligence in law- Improve Decision-Making.October 2023Artificial
making and law enforcement: Risks and new opportunities”. Intelligence, Learning from AI, Decision-
Vestnik of Saint Petersburg University. Law 2: 516–526. making,Professional Go players, AI and Inequality.
https://doi.org/10.21638/spbu14.2024.214. [5] Raj Kumar Sharma][2023] Exploring the Ethical Implications
[2] [Ankita Shukla, 2Dr. Alka Agnihotri][2024] AI-Driven of AI in Legal Decision-Making.Indian Journal of Law.Vol.1
Smart Management Processes:Transforming Decision- | Issue 1 | Nov – Dec 2023https://doi.org/10.36676/law.2023-
Making and Shaping the Future. Library Progress v1i1-06.
International, 44(3), 5471-5493.Ankita Shukla, Alka [6] [Mukhammad Ali Turdialiev] [2024] Navigating the Maze:
Agnihotri Library Progress International| Vol.44 No.3 | Jul- AI and Automated DecisionMaking Systems.International
Dec 2024 Journal of Law and Policy |Volume: 2, Issue: 7 .



Proceedings of WRFER International Conference, Pune, India, 28th June, 2025

14
 A CASE STUDY OF SYNTHESIS IN VLSI DESIGN USING FUSION
COMPILER
1
SAI RAJU NUGURI, 2RK SUNKARIA

Dept. of ECE, National Institute of Technology Jalandhar, Punjab

E-mail: [email protected], [email protected]

Abstract - Three main constraints—area, timing, and power consumption—have a major impact on the design flow of
Application Specific Integrated Circuits (ASICs) in the field of Very Large Scale Integration (VLSI) design. Addressing
these limitations becomes more crucial as these circuits' complexity and performance requirements increase. Electronic
Design Automation (EDA) tools, which facilitate several stages of the design process, have been developed to manage this
increasing complexity. Logic synthesis is a critical step that converts highlevel Register Transfer Level (RTL) code into an
efficient gatelevel netlist. To guarantee that the finished implementation satisfies predetermined performance criteria, this
transformation takes design limitations into account. Other factors are introduced when Design for Testability (DFT)
methodologies are used during synthesis. Although DFT makes the design more testable, it can also result in higher power
and area usage, which could affect performance as a whole. It is crucial to perform performance analysis, which includes
comparing the synthesis outcomes under different constraint conditions and with and without DFT insertion. These studies
aid in comprehending the trade-offs and in making well-informed judgments to successfully optimize the design. Netlists are
legally validated against RTL for Logic Equivalence Checking (LEC) using the Conformal EC tool. A Synopsys EDA tool
called Fusion Compiler (FC) offers a convenient way to combine synthesis approaches, reducing design cycles and
enhancing design quality.

Keywords - VLSI, Synthesis, Logic Equivalence Check(LEC), Fusion Compiler(FC).

I. INTRODUCTION and netlist is verified using the FEV approach. It is

now a crucial component of the design process and is
EDA tools are essential for the design, verification, used for hand editing later on, such as Engineering
and debugging of bigger digital circuits because of Change Orders (ECO). It is a comprehensive
the quick improvements in VLSI design level of verification technique that uses no test vectors to fully
integration during the past few decades. Designing a confirm design functioning. The powerful and
range of electronic chips using integrated circuits is effective instrument for logical equivalency is the
made possible by EDA tools. The scope for EDA Cadence Conformal Encounter Equivalence Checker.
tools is presented by the need to satisfy the functional The paper is structured as follows: 1. Creating an
and performance goals within a design time frame RTL model for the case study 2. Use the FC tool and
and associated cost. One such EDA tool is the perform Physical Synthesis. 3.Optimize the design for
Synopsys Fusion Compiler (FC), which combines best PPA. 4. Using a conformal tool, perform LEC of
user-defined gate arrays and standard cell libraries to the generated design. The proposed work's scope is
carry out the logic synthesis of intricate digital that new synthesis and optimization strategies that
circuits to create quick and space-efficient ASIC yield superior synthesis results are made possible by
designs. The Fusion compiler optimizes both the use of EDA tools such as design compilers.
sequential and combinatorial designs and supports a
broad variety of flat and hierarchical design patterns. II. DESIGN FLOW
Under varied loads, temperature, and voltage
circumstances, it also examines design trade-offs RTL modeling is the first step in the design flow
involving design restrictions like timing, area, and employed in the case study, and LEC is the last step.
power. Additionally, EDA tools are lowering Figure 1 provides an illustration of the design flow
manufacturing errors, cutting manufacturing costs, description.This process is dividedinto many sub-
optimizing IC design, and making complicated ICs steps, which include RTL Design,Verification,
easier to use. Meeting the targeted temporal and area synthesis, LEC. Figure 1 depicts the steps involved in
goals can improve performance while producing the integrated circuit (IC) design flow.
high-quality solutions. However, this presents a
difficulty for the logic synthesis of hierarchical Several essential documents are crucial in running
modular design utilizing the FC tool. Given a this flow.
standard-cell library and design restrictions, logic a) RTL File (all.v): The files provide a high-level
synthesis is the process of translating a high-level explanation of the digital circuit's functionality
design description into an optimal gate-level and behavior. They specify how the circuit
representation. It is a highly automated process that should behave and function.
connects physical design automation and high-level b) Technology Library Files (.lib): The timing,
synthesis. The logical equivalency between the RTL power, and area related to the standard cells are

Proceedings of WRFER International Conference, Pune, India, 28th June, 2025

15
 A Case Study of Synthesis in VLSI Design Using Fusion Compiler

represented in ASCII by the timing library (.lib). A. LOGIC SYNTHESIS

The output capacitance (Load) and input Given a standard-cell library and design restrictions,
transition (Slew) are used to calculate the delay. logic synthesis is the process of translating a high-
c) SDC: The timing and power-related limitations level design description into an optimal gate-level
that regulate design in relation to the representation. It is a highly automated process that
specification are included in standard design connects physical design automation and high-level
constraints, also known as Synopsys design synthesis. Logic synthesis encompasses a broad range
constraints. of digital system design, with synthesis approaches
d) UPF: An IEEE standard called the Unified primarily focusing on synchronous behavior and
Power Format (.upf) is used to specify power and implementation. The marketdriven choices that
associated elements of multi-voltage design. impact the design philosophies used for product
Definitions of supply sets, power domains, implementation have a significant impact on the
power switches, retention cells, level shifter quantity of design automation and logic synthesis
cells, and other low power-related terms are work. The RTL description in Verilog is first
included in UPF. converted by logic synthesis into components taken
e) lef file: The standard cells, macros, and other from the DesignWare and Generic Technology
design components needed for chip layout are (GTECH) libraries, which are independent of
described in depth. In addition to metal layer technology. After that, it was mapped to technology-
definitions and routing track specifications, it dependent library cells and optimized using the
contains details about the physical dimensions of appropriate optimization rules and methods.
cells, including height, width, and pin locations. Architectural optimization, logic level optimization,
The design guidelines for cell placement, routing and gate level optimization are the three layers of
restrictions, and via configurations are specified optimization that can be carried out during the
in the LEF file. synthesis process.
f) def file: It records the specific physical layout
details of an integrated circuit, such as the B.PHYSICAL SYNTHESIS
positioning of I/O pins, standard cells, and Following logic synthesis, an early design floorplan is
macros, as well as the interconnect routing. It created for placement data, and based on this
gives the coordinates for every cell and macro on floorplan, routing requirements are estimated. This
the chip, as well as the positions and orientations information is provided as an input in .def file to
of every component. The precise routes of the physical synthesis. It captures the detailed physical
interconnects as well as details on the metal layout information of an integrated circuit, including
layers and routing vias are contained in the DEF the standard cells, macros, and I/O pins placement, as
file. In order to guarantee that the design well as the routing of interconnects We are restricting
complies with the given design guidelines and the tool to place within a certain area. Real Physical
limitations, it also specifies the positioning of Synthesis is more accurate , but runtime is more.
blockages and routing impediments.
C.DFT Insertion
These days, DFT is crucial to the design flow of
VLSI ASICs since it makes it possible to include
testability characteristics in the design stage itself. It
guarantees that every flaw in the design will be
found. Adding additional signals and logic to a test in
the design is the underlying idea of structured DFT,
which is simple to implement. SCAN and BIST are
the two structured approaches that are most
frequently utilized. Many sequential elements in the
design can be made more controllable and observable
by using the SCAN Shift and Scan Capture approach.
The most well-liked and extensively applied
structured DFT method is SCAN Shift and Scan
Capture.
To make the process of creating test patterns easier,
the scan approach converts all sequential elements
into combinational elements. By joining all of the flip
flops in series to create a shift register known as Scan
Chains (SC), the scan technique aims to incorporate
the test mode logic into a circuit. It links scan inputs,
or PIs, and scan outputs, or POs.
Figure.1: Design flow for case study

Proceedings of WRFER International Conference, Pune, India, 28th June, 2025

16
 A Case Study of Synthesis in VLSI Design Using Fusion Compiler

All of the original conventional flip flops in the non-equivalents, the tool employs gate reporting.
design can be swapped out for scan flip flops (SFFs) Visual accessibility is enhanced by the schematic
to create a scan chain. Three different kinds of scan representation and color coding of error candidates,
flip flips, including LSSD, multiplexed, and clocked nonequivalent points, diagnostic input points, and
forms, can be employed in scan insertion. supporting critical points.
Multiplexed scan cells are the most popular flip flop
in scan design because they work well with III. IMPLEMENTATION
contemporary designs and are fully supported by
automation technologies now in use. A flip flop's Verilog HDL is utilized to explain the case study's
input is coupled to two more input multiplexers in a RTL design. Writing the RTL verilog code in a way
multiplexed SFF. that facilitates synthesis is essential.The Fusion
Compiler tool accepts the case study's RTL verilog
D. Logic Equivalence Check code as input. Following the HDL to gate translation,
The logic equivalency between the netlist and RTL is the FC tool optimizes the design and maps it to a
verified using LEC. It is now a crucial component of particular technology library known as Target
the design process and is used for hand editing later Library. This process is guided by constraints. Details
on, such as Engineering Change Orders (ECO). It is a on cell characteristics, features, and design operating
comprehensive verification technique that uses no test conditions are included in the target library. Using the
vectors to fully confirm design functioning. If both symbol library, the synthesis tool creates a gate level
models are logically identical, LEC is a very quick schematic and an RTL design schematic. To resolve
and simple procedure that essentially confirms the references, the fc tool makes use of both the link and
rapid modifications to the model designs. target libraries. All libraries with cells referred by the
The golden model is the name given to the RTL netlist are listed in the link library. Time and path
design paradigm at the behavioral abstraction level. delays are computed using delay models defined by
This is combined to create the netlist, which is the link library. The designer has additional control
regarded as the updated model at the gate level of over how the cells are mapped thanks to the target
abstraction. There are two LEC modes: setup and and link libraries. They can also remap a gate-level
LEC. The RTL design's constraints are specified by netlist between different technologies.
the setup mode. The LEC technique uses the The analysis command transforms HDL designs into
mapping, comparison, and diagnosis procedures to an intermediate format in the work directory after
validate the RTL design in LEC mode. In the LEC examining them for syntax and synthesizable logic.
mode, it first maps each significant design point. The Elaborate then takes this as input and starts working
necessary parts include tie z, tie gates, black boxes, on turning the RTL description into physical
cut gates, D latches, D flip flops, primary inputs, and hardware. The processes of analysis and elaboration
primary outputs. Combinational logic cones enclose quickly determine whether the design intent of the
the design's primary points. The cones of logic must tools is being met. During synthesis, the
line up in order to compare two designs using the key check_design command is used to fix design flaws
point mapping process. and verify the internal representation. Unconnected
ports, constant valued ports, recursive hierarchy
 Mapping definitions, connection class violations, cell and
For later comparison, mapping compares important reference mismatches, numerous driver nets, and cells
elements of the original design with their equivalents lacking input or output pins are among the errors that
in the modified design. Function-based mapping and occur during check_design.
name-based mapping are the two categories of key The FC tool reads optimization constraints as input.
point mapping. Module and signal names are mapped The Fusion compiler optimizes designs using design
in designs using name-based mapping. It takes longer rule restrictions and optimization constraints during
to map logic functions like gates, flops, and primary synthesis. Technology libraries define implicit design
inputs in models using function-based mapping than rules. These requirements are necessary for a design
name-based mapping. to work properly and can be stricter than the
 Compare optimization constraints. It includes the maximum
Use the comparison command in LEC mode after fan-out, the minimum and maximum capacitances,
mapping important locations with the conformal tool. cell degradation, and the maximum transition time.
Black boxes, D flip-flops, D latches, and principal These properties cannot be removed because they are
outputs are examples of comparison points that essential for technology, however more limited
designers can manually define. values can be set.
 Analyze The design's speed and area objectives are
The method looks for possible faults by examining represented by the optimization constraints. Area is
non-equivalent points. The conformal tool shows the not as crucial as speed. Design rule constraints take
underlying error patterns and detects non-equivalent precedence over these limits. Timing limitations for
spots. To determine the fan-in or fan-out cones of synchronous pathways include input and output

Proceedings of WRFER International Conference, Pune, India, 28th June, 2025

17
 A Case Study of Synthesis in VLSI Design Using Fusion Compiler

delays, while asynchronous paths have maximum and manager, schematic tracer, source code manager, and
minimum delays. Timing for critical paths, false gate manager.
paths, and multi-cycle paths should be carefully
considered. The gui_start command loads the FC IV. RESULTS AND DISCUSSIONS
tool's design vision, which allows RTL and netlist
schematics to be shown. It provides dialogue boxes 1. Timing Report
for commands to the FC. It also aids in analyzing the After executing the synthesis, we observed setup
design at the GTECH and gate levels. violations in the design. Achieving timing clean
The LEC tool has two modes: setup and LEC mode. results in a design that meets all timing specifications
To use the LEC tool in setup mode, designers must , ensuring functional correctness and reliability.
input the golden RTL and revised gate level designs,
as well as the standard libraries needed for DC
synthesis. The design attributes and constraints are
also defined, which are required for the LEC tool to
perform equivalency checks on the designs. Table.1: Timing Report after the Synthesis execution
The designer can define comparison points between
golden and amended designs. The DFF and POs are NUM – No of Violating Paths
defined as comparison points. If the compare points WNS – worst negative Slack
are comparable, the RTL and netlist designs are TNS – total negative Slack
logically equivalent. To diagnose unmapped or non- Figure 3 illustrates the flow of formal equivalence
equivalent sites, use tools such as the diagnosis verification using the cadence conformal tool.

2. Strategy used for fixing Timing Violation timing is clean, with reduced power consumption and
In the first step initially focused on the critical paths minimal increase in area.
which are causing setup violations. Generally setup
violations occur when your data path is slower. So I
made data path faster by including ULVT cells, for
which delay will be less.
Table.3: Timing Report after second iteration

3. Power Report

Table.2: Timing Report after first iteration

After this power has been increased because of

ULVT cells.In the second iteration I have increased Table.4: Power Report
the drive strength of the cells in the critical paths and
reduced the percentage of ULVT cells. After the next After synthesis, the Conformal EC tool is used to run
iteration there is no new timing violation and the the LEC with inputs including golden and amended
Proceedings of WRFER International Conference, Pune, India, 28th June, 2025
18
 A Case Study of Synthesis in VLSI Design Using Fusion Compiler

netlist designs, as well as the standard cell library. Circuits and Systems, vol. 39, no. 12, pp. 5111-5122, Dec. 2020,
doi: 10.1109/TCAD.2020.2966551.
The LEC run is successful if both designs have an [2] S. H. Loh, I. M. Tan and J. J. Sim, "VLSI Design Course with
equal number of mappings and compare important Commercial EDA Tools to Meet Industry Demand – From Logic
points. The golden and updated models each have Synthesis to Physical Design," 2021 11th IEEE International
Conference on Control System, Computing and Engineering
18268 critical locations for mapping logical (ICCSCE), Penang, Malaysia, 2021, pp. 55-60, doi:
equivalence checks. The LEC does not use PI as a 10.1109/ICCSCE52189.2021.9530746.
crucial factor to compare both designs. The result of [3] Attaoui, Y., Chentouf, M., Ismaili, Z. E. A. A., & El Mourabit, A.
(2021, December). Clock Gating Efficiency and Impact on Power
LEC run is summarized in Table 5. Optimization During Synthesis Flow. In 2021 International
Conference on Microelectronics (ICM) (pp. 13- 16). IEEE.
[4] S. Gayathri and T. C. Taranath, "RTL synthesis of case study
using design compiler," International Conference on Electrical,
Electronics, Communication, Computer, and Optimization
Techniques (ICEECCOT), Mysuru, India, 2017, pp. 1-7, doi:
10.1109/ICEECCOT.2017.8284603.
[5] Goudar, S., & Aradhya, H. R. (2020, July). Development and
Analysis of Parameters to Evaluate Design Partitioning of SoC.
Table.5: Mapped Points of LEC
In 2020 Second International Conference on Inventive Research
in Computing Applications (ICIRCA) (pp. 416-421). IEEE.
[6] A. Agnesina, K. Chang and S. K. Lim, "Parameter Optimization
of VLSI Placement Through Deep Reinforcement Learning," in
IEEE Transactions on Computer-Aided Design of Integrated
Circuits and Systems, vol. 42, no. 4, pp. 1295-1308, April 2023,
doi: 10.1109/TCAD.2022.3193647.
[7] M. I. Niranjana, J. Dhanasekar, N. Blesson, G. D. Immansingh,
M. Boopathi and S. M. Raman, "A Framework for Block-Level
Physical Design using ICC2 in 14nm Technology," 2023 7th
International Conference on Electronics, Communication and
Aerospace Technology (ICECA), Coimbatore, India, 2023, pp.
413 417, doi: 10.1109/ICECA58529.2023.10394906.
[8] W. A. Pleskacz, C. H. Ouyang and W. Maly, "A DRC based
algorithm for extraction of critical areas for opens in large VLSI
circuits," in IEEE Transactions on Computer-Aided Design of
Integrated Circuits and Systems, vol. 18, no. 2, pp. 151-162, Feb.
Figure 4: Comparision of LEC 1999, doi: 10.1109/43.743724.
[9] Y. -C. Li, S. -Y. Lin, H. -Y. Lin and J. C. -M. Li, "Diagnosis and
repair of cells (DRC) responsible for power-supply-noise
V. CONCLUSIONS violations," 2018 International Symposium on VLSI Design,
Automation and Test (VLSI-DAT), Hsinchu, Taiwan, 2018, pp.
Using EDA tools throughout the design process 1-4, doi:
[10] 10.1109/VLSI-DAT.2018.8373274. 10. Y. -Y. Huang, C. -T. Lin,
improves accuracy and quality of outcomes. W. -L. Liang and H. -M. Chen, "Learning Based Placement
Designing using hierarchy, splitting, and parallelism Refinement to Reduce DRC Short Violations," 2021 International
reduces operation time. Using EDA tools can save Symposium on VLSI Design, Automation and Test (VLSI-DAT),
Hsinchu, Taiwan, 2021, pp. 1-4, doi: 10.1109/VLSI
design time and costs. The design compiler uses its DAT52063.2021.9427321.
builtin mapping and optimization tools to optimize [11] Andrew B. Kahng, Jens Lienig, Igor L. Markov, Jin Hu, ―VLSI
the case study. The example study effectively Physical Design: From Graph Partitioning to Timing Closure‖,
First Edition, Springer Dordrecht 2011.
implements most of the ASIC design flows, including [12] J. Chen and W. Zhu, "An Analytical Placer for VLSI Standard
RTL model development and LEC. Our work offers Cell Placement," in IEEE Transactions on Computer-Aided
significant insights and practical solutions for VLSI Design of Integrated Circuits and Systems, vol. 31, no. 8, pp.
1208-1221, Aug. 2012, doi: 10.1109/TCAD.2012.2190289.
designers, helping to optimize verification processes [13] W. -T. Hung, Y. -G. Chen, J. -G. Lin, Y. -W. Yang, C. H. Tsai
in the semiconductor industry. Our technique aims to and M. C. -T. Chao, "DRC Violation Prediction After Global
improve the efficiency and reliability of integrated Route Through Convolutional Neural Network," in IEEE
Transactions on Very Large Scale Integration (VLSI) Systems,
circuits, which are essential for modern electronics. vol. 31, no. 9, pp. 1425 1438, Sept. 2023, doi:
10.1109/TVLSI.2023.3271932.
Future work will focus on further improving the [14] S. Raman and M. Lubyanitsky, "Cone Resynthesis ECO
Methodology for Multi-Million Gate Designs," 2009 22nd
optimization techniques and expand their application International Conference on VLSI Design, New Delhi, India,
to other VLSI blocks and technologies. Our research 2009, pp. 195-199, doi: 10.1109/VLSI.Design.2009.28.
highlights the significance of novel methodologies in [15] Jovanović, B., & Jevtić, M. (2012). Methods for power
minimisation in modern VLSI circuits. International Journal of
VLSI design, resulting in more dependable and high- Reasoning-based Intelligent Systems, 4(1-2), 46-57. 16. T. -W.
performance semiconductor devices. Tseng, C. -T. Lin, C. -H. Lee, Y. -F. Chou and D. -M. Kwai, "A
power delivery network (PDN) engineering change order (ECO)
approach for repairing IR-drop failures after the routing stage,"
REFERENCES Technical Papers of 2014 International Symposium on VLSI
Design, Automation and Test, Hsinchu, Taiwan, 2014, pp. 1-4,
[1] M. I. Dewan and D. H. Kim, "NP-Separate: A New VLSI Design doi: 10.1109/VLSI-DAT.2014.6834874.
Methodology for Area, Power, and Performance Optimization,"
in IEEE Transactions on Computer-Aided Design of Integrated



Proceedings of WRFER International Conference, Pune, India, 28th June, 2025

19
 PERFORMANCE ANALYSIS OF DOUBLE-GATE STAGGERED
HETEROJUNCTION TFET BIOSENSOR
1
RUPALI, 2DR B S SAINI

Department of Electronics & Communication Engg., NIT Jalandhar Punjab 144011, India
E-mail: [email protected], [email protected]

Abstract - This study introduces a double-gate staggered heterojunction tunnel field-effect transistor (DG-SHJ-TFET)
optimized for high-precision biosensing. The design features an InAs heterojunction pocket at the source-channel interface
and a nanoscale cavity on the source side, which collectively improve band-to-band tunneling efficiency and overall sensor
sensitivity.The study investigates the impact of biomolecule dielectric constants (εk = 1 to 12) and charge densities (neutral,
positive, negative) on device electrostatics, including surface potential, energy band alignment, and drain current
characteristics. Results demonstrate that increasing the dielectric constant (εk) improves sensitivity, with positively charged
biomolecules exhibiting the highest response due to enhanced electric field effects. The device achieves 60× higher drain
current sensitivity (SId), almost double the threshold voltage shift (ΔVth), and 12 times increase in ON-OFF current ratio
sensitivity (Δ(Ion/Ioff)) compared to conventional FET-based biosensors, further optimizing performance. These findings
highlight the potential as a highly sensitive, low-power biosensor for detecting biomolecules in medical diagnostics,
environmental monitoring, and lab-on-chip applications.

Keywords - DG-SHJ-TFET, Biosensor, Sensitivity

I. INTRODUCTION dielectric constant or the charge content of the

molecule results in effective variation of electrical
The continuous advancement in nanotechnology and properties of the sensor. The proposed structure is
biomedical engineering has accelerated the simulated and analyzed thoroughly, along with
development of miniaturized, low-power biosensors discussing its performance with the crucial
capable of detecting biomolecules with high parameters of biosensing like sensitivity, Ion/Ioff
sensitivity and specificity. Some of the transistor ratio, and subthreshold swing, which demonstrates its
structures include Tunnel Field-Effect Transistors. suitability in the future biological sensing
TFETs have also demonstrated potential in technologies [5].
biosensing applications, taking advantage of their
lowvoltage operation and sub-thermal subthreshold II. DEVICE ARCHITECTURE AND
swings. SIMULATION SETUP

The traditional MOSFET-based biosensors have the The biosensor incorporates a double-gate TFET
weaknesses of excessive leakage currents and poor device structure that has a strategically engineered
sensitivity at the nanoscale size [1-2]. Instead, band- hetero-pocket formed at the source-channel junction
to-band tunnelling (BTBT) is the dominant carrier that significantly increases the efficacy of band-to-
injection mechanism employed by TFETs, and band tunneling (BTBT). The device is designed and
TFETs rely on it to produce a steep subthreshold manufactured on 100 nm silicon substrate making the
slope and very low off-state current. Because of these best use of its compact structure to provide better
properties, TFETs are well suited to applications performance. One of the key constituents is a 10 nm
where low power and high sensitivity are required, thick layer of HfO2 deposited by LPCVD and an
e.g., handheld diagnostic and integrated lab-on-chip aluminum gate etched in dry etching. The selective
platforms. etching of HfO2 layer is done to form nanocavities on
both sides of the gate close to the source area.
Other structural developments such as many gate
topologies, and staggered heterojunctions, have also A tiny layer of silicon is exposed by the nano-
been explored to enhance the performance of TFET cavities, which immobilizes biomolecules. Charge
biosensors [3]. The staggered heterojunction provides transport is enhanced by electrodes around the source
better tunnelling characteristics by tuning the energy and drain and minimized leakage. Simulations were
band alignment between the source and the channel carried out using Silvaco TCAD tools, utilizing
regions and the stacked gate structure provides better realistic models like carrier mobility, recombination,
electrostatic control over the channel that results in and quantum transport. incorporating a concentration-
improved switching characteristics. In this thesis, a dependent mobility model, a recombination model
new kind of biosensor is studied using a Double Gate used in this like consrh, and a Fermi- Dirac statistics
Staggered Heterojunction TFET (DG-SHJ-TFET) [4]. transport model to account for the effects of heavy
A nanogap is created in the gate dielectric area to doping densities. Table I shown the device
interact with the target biomolecules. A change in the specification.

Proceedings of WRFER International Conference, Pune, India, 28th June, 2025

20
 Performance Analysis of Double-Gate Staggered Heterojunction TFET Biosensor

III. RESULTS AND DISCUSSIONS

A. DC Characteristics
The DC characteristics of the proposed DG-SHJ-
TFET biosensor were examined using TCAD
simulations to understand its behaviour in various
situations. For εk = 1 (air), the device shows minimal
sensitivity, but as εk increases (3, 5, 8, 10, 12), the
drain current sensitivity (SId) improves substantially.
These properties describe the connections between
the gate voltage that is applied and the consequent
drain current and are significant to evaluate the
functioning of sensors. Such parameters as energy
band diagrams, electric field distribution, tunnelling
rate, and Id- Vgs behaviour were studied with
Fig.1 Proposed Device Structure emphasis on the role of interface traps, gate oxide
properties, and biomolecular charges. [6-7].

Table I: Device specification

Fig.3 DC characteristics for Neutral Biomolecules (a) Energy-
Band Diagram, (b) Id vs Vgs Graph, (c) Electric Field
Assurance of calibration against experimental results Variation
accurate prediction of the biosensor's electrical
behavior. This design provides encouraging Variations in ON current, subthreshold swing (SS),
advantages for label-free detection of biomolecules at and threshold voltage were analyzed to understand
the nanoscale [9]. the sensitivity and accuracy of the biosensor. The
implemented PNPN configuration in the TFET
architecture optimizes charge carrier tunneling
through strategic electric field modulation at the
source-channel interface [8]. This structural
enhancement yields superior ON-state characteristics
and achieves a more abrupt subthreshold swing (SS).
Furthermore, the inclusion of a heavily doped n-type
pocket between source and channel regions generates
intensified electric field concentrations at the p-n
junction boundary. This dual effect simultaneously
amplifies current conduction while effectively
mitigating interface trap formation [10].

B. Effect of Positively Charged Biomolecules

For εk = 1 (air), the device shows minimal sensitivity,
Fig.2 Calibration graph but as εk increases (3, 5, 8, 10, 12), the drain current

Proceedings of WRFER International Conference, Pune, India, 28th June, 2025

21
 Performance Analysis of Double-Gate Staggered Heterojunction TFET Biosensor

sensitivity (SId) improves substantially. At εk = 12,

the highest sensitivity is observed due to the
combined effect of the dielectric constant and
positive charge concentration. The positive charges
change the electric field and width of the depletion
region therefore, the threshold voltage sensitivity
(ΔVth) also increases, as the electric field and
depletion region width are altered by the positive
charges.

These results prove the functionality of the sensor in

the precise detection of the presence of positively
charged biomolecules. Simulation with positively
charged biomolecules showed enhanced channel
surface potential, which intensified the electric field
at the tunneling junction. A fixed positive charge
density of 1 × 10¹¹ cm⁻² was introduced for this
analysis. The stronger electric field leads to improved
band-to-band tunneling, raising the drain current. The Fig.5 (a) DC characteristics for Negative Biomolecules (a)
band diagram also revealed an upward bending near Energy-Band Diagram, (b) Id vs Vgs Graph, (c) Electric Field
the channel surface, which enhances the tunneling Variation
current. Therefore, the presence of positively charged
biomolecules significantly improves the biosensor‟s When positively charged biomolecules (Nbio = 1 ×
sensitivity by increasing the band bending and 10¹¹ cm⁻²) are introduced into the nano-cavity, the
tunneling efficiency [11-12]. device exhibits significant improvements in
sensitivity [15]. The presence of positive charges
elevates the surface potential at the tunnelling
interface, intensifying the local electric field
magnitude. This field enhancement enhances electron
tunneling efficiency, thereby increasing the drain
current (Id). The mechanism occurs through effective
gate voltage amplification, which enhances charge
carrier accumulation in the conduction channel [16].
The energy band diagram shows an upward shift in
the conduction and valence bands, reducing the
tunneling barrier and enhancing current flow. In
Table II shown biomolecules used in this paper for
analysis and comparison.

Fig. 4 DC characteristics for Positive Biomolecules (a) Energy-

Band Diagram, (b) Id vs Vgs Graph, (c) Electric Field
Variation

C. Effect of Negatively Charged Biomolecules

In contrast, negatively charged biomolecules were Table II: Biomolecules and Their Dielectric Constant
found to suppress the tunneling process. These
charges reduce the surface potential, weakening the IV. SENSITIVITY ANALYSIS
electric field at the junction. Consequently, the
tunneling probability decreases, leading to a Two significant figures were used to characterize the
noticeable reduction in drain current [13]. The energy sensitivity of the DG-SHJ-TFET used to assess the
band diagram shifts downward, increasing the barrier biosensor: drain current sensitivity (SId), threshold
height for tunneling. This drop in current response voltage shift (ΔVth), and ON-OFF current ratio
confirms the biosensor‟s ability to distinguish sensitivity (Δ(Ion/Ioff)). Such parameters were
between biomolecule polarities, with reduced considered both in neutral and charged biomolecules
sensitivity in the case of negative charges [14]. in different conditions of dielectric constant.

Proceedings of WRFER International Conference, Pune, India, 28th June, 2025

22
 Performance Analysis of Double-Gate Staggered Heterojunction TFET Biosensor

Fig. 8 Sensitivity Analysis for Negative Charge Biomolecules

A. Sensitivity Analysis for Neutral Bio-Molecules (a) Ion/Ioff sensitivity, (b) Drain current (Id) sensitivity
Neutral biomolecules (Nbio = 0) do not contribute
additional charges but influence the device through For biomolecules carrying negative charges, the
dielectric modulation. As εk increases (1 to 12), the sensor exhibits reduced sensitivity; however, the
effective gate capacitance changes, modifying the response remains detectable. Fig. 8 shown that the
tunneling junction's electric field. The drain current sensitivity analysis of positive charge biomolecules.
sensitivity (SId) improves with higher εk, as the At εk = 12 and Nbio = -1 × 10¹¹ cm⁻², the drain
dielectric environment enhances carrier tunneling. current sensitivity is reduced due to the opposing
The threshold voltage sensitivity (ΔVth) also exhibits effect of negative charges on the electric field.
a steady rise, demonstrating the device's capacity to However, the device remains responsive, highlighting
identify neutral biomolecules by their dielectric its versatility in detecting varying biomolecule types
characteristics. However, the sensitivity is less than [19-22]. For comparative analysis taken state of arts
that of charged biomolecules due to the absence of biosensors. In Table III shown this work comparison
charge-induced band bending [17-18]. with recants biosensors.

Fig. 6 Sensitivity Analysis for Neutral Biomolecules (a) Ion/Ioff

sensitivity, (b) Drain current (Id) sensitivity

B. Sensitivity Analysis for Charged Bio-Molecules Table III Sensitivity Comparison of this Biosensor with the
For positively charged biomolecules, the sensitivity State-of-the-art Biosensors
metrics (SId, Δ(Ion/Ioff)) exhibit significant
enhancements with increasing εk and charge V. CONCLUSIONS
concentration. At εk = 12 and Nbio = 1 × 10¹² cm⁻²,
the drain current sensitivity improves by 60 times, This research introduces a novel dielectric-modulated
while the threshold voltage sensitivity doubles biosensing platform based on an n-pocket double gate
compared to conventional FET-based biosensors. The staggered heterojunction tunnel field-effect transistor
ONOFF current ratio sensitivity increases 12-fold, (DGSHJ- TFET) architecture. The proposed device
demonstrating the device's superior performance. In demonstrates enhanced sensitivity for biomolecular
fig.7 shown that the sensitivity analysis of the detection through its unique charge-modulation
negative charge biomolecules. capabilities. The device leverages dielectric
modulation and charge-dependent sensitivity
variations to achieve superior performance compared
to conventional FET-based biosensors. The source-
channel junction InAs pocket helps to increase the
tunnelling efficiency, and the nano-cavity on the
source side at both gate allows the controlled
biomolecule immobilization. Important results
include the fact that sensitivity is highly improved
with the dielectric constant (εk = 1 to 12) and that
Fig. 7 Sensitivity Analysis for Negative Charge Biomolecules positively charged biomolecules give the largest
(a) Ion/Ioff sensitivity, (b) Drain current (Id) sensitivity response because the electric field effects are more
pronounced. The biosensor has a doubling threshold

Proceedings of WRFER International Conference, Pune, India, 28th June, 2025

23
 Performance Analysis of Double-Gate Staggered Heterojunction TFET Biosensor

voltage sensitivity (ΔVth), a 60-fold increase in drain [11] N. N. Reddy and D. K. Panda, “A Comprehensive Review on
Tunnel Field-Effect Transistor (TFET) based biosensors:
current sensitivity (SId), and a 12 times improvement recent advances and future prospects on device structure and
in ON-OFF current ratio sensitivity (Δ(Ion/Ioff)). sensitivity.” Silicon, 13, 3085 (2020).
These improvements are owed to an increased [12] R. Ghosh, R. P. Nelapati, P. Saha, R. Chinthaginjala, T. Kim,
efficiency of band-to-band tunneling (BTBT), and K. S, “Sensitivity analysis of bi-stacked gateoxide hetero-
junction tunnel fet with Si0.6Ge0.4 source biosensor
reduced leakage currents, and an adequate considering non-ideal factors.” PLOS ONE, 19, e0301479
suppression of shortchannel effects (SCEs). In (2024).
summary, the proposed Double Gate Staggered-TFET [13] S. K. Verma, S. Singh, G. Wadhwa, and B. Raj, “Detection of
biosensor offers high sensitivity, low power biomolecules using Charge-Plasma based gate underlap
dielectric modulated doping less TFET.” IEEE Transactions
consumption, and scalable fabrication, making it a on Electrical and Electronic Materials, 21, 528 (2020).
promising candidate for label-free, real-time [14] L. Li, C. Li, Z. Zhang, and E. Alexov, “On the dielectric
biomolecule detection in medical diagnostics, „Constant‟ of proteins: smooth dielectric function for
environmental monitoring, and point-of-care testing. macromolecular modeling and its implementation in DelPhi.”
Journal of Chemical Theory and Computation, 9, 2126
Future efforts will aim at validating the design (2013).
experimentally and refining it for targeted [15] J. Talukdar, G. Rawat, and K. Mummaneni, “Dielectrically
biomolecule detection, with the goal of improving its modulated single and double gate tunnel FET based
effectiveness and potential for real-world biosensors for enhanced sensitivity.” IEEE Sens. J., 21,
26566 (2021).
applications. [16] S. Anand, A. Singh, S. I. Amin, and A. S. Thool, “Design and
performance analysis of dielectrically modulated doping-less
REFERENCES tunnel FET-based label free biosensor.” IEEE Sensors J., 19,
4369 (2019).
[1] M. G. Virak and R. Chaujar, “TCAD analysis and modeling [17] V. Thakur, A. Kumar, and S. Kale, “Analytical modelling of
of gatestack gate all around junctionless silicon nanowire space engineered reconfigurable silicon nanowire Schottky
TFET as label-free bio-sensor for biomedical application.” barrier transistor for biosensing applications.” Micro
Proc. IEEE VLSI Device Circuit Syst., 60 (2020). Nanostruct., 188, 207795 (2024).
[2] K. Vanlalvampuia and B. Bhowmick, “Analysis of [18] H. Xie, H. Liu, S. Chen, T. Han, and S. Wang, “Electrical
heterostacked source TFET and heterostructure vertical FET performance of InAs/ GaAs0.15In0.9 heterostructure
as dielectrically modulated label-free biosensors.” IEEE Sens. junctionless TFET with dual-material gate and
J., 22, 939 (2022). Gaussiandoped source.” Semicond. Sci. Technol., 35, 095004
[3] K. Ganapathi, Y. Yoon, and S. Salahuddin, “Analysis of InAs (2020).
vertical and lateral band-to-band tunnelling transistors: [19] K. R. Singh, A. Naugrahiy, and G. P. Mishra, “Endurance
Leveraging vertical tunneling for improved performance.” behavior of Z-shaped charge plasma tunnel FET for
Appl. Phys. Lett., 97, 033504 (2010). biosensing application.” Journal of Circuits, Systems and
[4] A.Rasoool, S. Kossar, S. Parveen, and U. Rasool, Computers, 33 (2024).
“Heterojunction Tunnel Field-Effect Transistors (TFETs) and [20] A. Kumar, Y. Thakur, S. Kumar, S. Kale, and K. R. Singh,
applications.” Handbook of Emerging Materials for “Sensitivity investigation of underlap gate cavity-based
Semiconductor Industry 471 (2024). reconfigurable silicon nanowire schottky barrier transistor for
[5] R. Meshkin, “An extended-source tunneling-FET with gate- biosensor application.” Silicon, 16, 5877 (2024).
overlapped n+-doped pocket.” J. Electron. Mater., 52, 5980 [21] M. R. Tripathy, A. Kumar Singh, S. Chander, P. K. Singh, K.
(2023). Baral, and S. Jit, Device-Level performance comparison of
[6] A.Kumar and S. Kale, “Numerical modeling and performance some pocket engineered III-V/Si heterojunction vertical
analysis of underlap gate cavity integrated reconfigurable tunnel field effect transistor, 2020 5th International
silicon nanowire Schottky barrier transistor biosensors.” Conference on Devices, Circuits and Systems (ICDCS),
Appl. Phys. A, 1 (2024). Coimbatore, India, 180 (2020).
[7] A.Kumar and S. Kale, “Analytical modeling of silicon [22] A. Swati, J. Kaur, and A. K. Singh, “Performance
nanowire dielectric modulated schottky barrier reconfigurable investigation of Ge-based dielectric modulated junctionless
FET (SiNW-DM SB RFET) biosensor.” ECS J. Solid State TFET as a label-free biosensor.” Appl. Phys. A, 130 (2024).
Sci. Technol., 13, 113005 (2024). [23] S. Singh, A. Naugrahiy, and G. P. Mishra, “Endurance
[8] A. Kumar and S. Kale, “Spacer-engineered reconfigurable behavior of Z-shaped charge plasma tunnel FET for
silicon nanowire schottky barrier transistor as a label-free biosensing application.” Journal of Circuits, Systems and
biosensor.” Silicon, 16, 723 (2024). Computers, 33 (2024).
[9] F. Mayer et al., “Impact of SOI, Si1-xGexOI and GeOI [24] M. Vadizadeh, “Digital performance assessment of the dual-
substrates on CMOS compatible tunnel FET performance.” material gate gaas/inas/ge junctionless TFET,” IEEE
IEMD Tech. Dig., Dec., 1 (2008). Transactions on Electron Devices, vol. 68, pp 1986–1991,
[10] S. R. Upadhyay, S. Kale, and A. Pandey, “Dielectric Feb 11 2021.
Modulated InAs Pocket Heterojunction Tunnel FET for [25] S. Kumar, Y. Singh, B. Singh, and P. K. Tiwari, “Simulation
Biosensor Applications,” ECS Journal of Solid State Science study of dielectric modulated dual channel trench gate TFET-
and Technology, vol. 14, no. 4, p. 047006, Apr. 2025, doi: based biosensor,” IEEE Sensors Journal,vol 20, pp. 12565–
https://doi.org/10.1149/2162-8777/adc338. 12573, Nov 2020..



Proceedings of WRFER International Conference, Pune, India, 28th June, 2025

24
 MEDI MOLECULE: AN AI-POWERED PLATFORM FOR
ACCELERATING DRUG DISCOVERY THROUGH MOLECULE
GENERATION AND REAL-TIME COLLABORATION
1
SOLAPURAM SHREEYA, 2DR. R. M. NOORULLAH, 3ANNALDAS SHIVANI
1
B. Tech., Student, 22951A05K4 Department of CSE Institute of Aeronautical Engineering Hyderabad, India
2
Associate Professor Department of CSE Institute of Aeronautical Engineering Hyderabad, India
3
B. Tech., Student, 22951A05K0 Department of CSE Institute of Aeronautical Engineering Hyderabad, India
E-mail: [email protected], [email protected], [email protected]

Abstract - MediMolecule is transforming the drug discovery scene with its AI-driven platform that addresses the usual
challenges of pharmaceutical research. Conventional drug discovery processes can be time-consuming and labor- intensive,
usually being hindered by the manual analysis of data. With the use of advanced AI models, such as Nvidia MolMIM, Medi
Molecule improves searching for, creating, and analyzing molecules. With its strong integration with databases such as Pub
Chem and RD Kit, the platform provides researchers with large chemical data and strong molecular modeling capabilities. It
also has real-time collaboration features that en able smooth interaction and sharing of ideas by researchers. Interactive data
visualizations offer rich, actionable insights into molecular properties, allowing scientists to make informed decisions.
MediMolecule is built with scalability, security, and usability in mind, making it a vital tool for scientists, academia, and
pharmaceutical firms to utilize. This groundbreaking platform showcases the tremendous power that AI technologies have to
contribute to drug discovery, significantly reducing time and expense while promoting increasedaccessibilityandinnovationin
developing life-saving drugs.

Keywords – Molecular Modeling, Nvidia Mo lMIM, Pub Chem, Pharmaceutical Research, Real-Time Communication, RD
Kit.

I. INTRODUCTION incorporation of data visualization technologies that

permit researchers to more carefully scrutinize and
Traditionally, finding medications takes time. Its the evaluate difficult molecules informations.
resource-intensive project, often further made MediMolecule offers visualizations and interactive
difficult bythe great amount of data to be sorted and maps for users. Offers significant observations
the difficulty of finding several promising drug regarding molecular qualities, associations, and
candidates. Mostly based on trial and error, constructed works. The presence of this property
traditional approaches to drug identification is helps scientists to recognize more straightforwardly
experiential testing, human analysis of data elements patterns and links sometimes otherwise overlooked.
that can add cost and lengthen schedules. In recent Adopting this data-driven approach helps scientists
times, though, the development of artificial make more reasoned judgments as molecular
intelligence (AI) and data-driven technologies has evidence is presented in more understandable and
started to redefine this space. AI-enabled tools offer a simpler form, hence allowing then concentrate on the
promising prospect to accelerate drug discovery most hopeful pharmaceuticals candidates.
through the automation of data analysis, optimization
of molecular design, and enabling better collaboration II. RELATED WORK
in research.
Over the past few years, advances in artificial
MediMolecule differentiates itself through its intelligence (AI) have immensely contributed towards
sophisticated molecule generation function, using drug research, particularly in molecular drug
Nvidia's Mol MIM model. This cutting-edge AI discovery and generation. The union of AI and
model enables the production of new molecules with pharmaceutical sciences has resulted in revolutionary
specific properties, essentially going beyond trial- findings which have accelerated the drug
and-error techniques. Through machine learning, development process, allowing researchers to
MediMolecule can predict how molecules will act discover new molecular shapes at a previously
with high accuracy, thus enabling researchers to find unimagined rate [1]. Deep learning algorithms have
promising candidates for drugs more quickly. This enabled novel molecular structure designs with
strategy, which utilizes AI, not only accelerates the improved features, thus reducing the time and
process of drug discovery but also opens the way for expenditure related to conventional drug
revealing molecular characteristics that have perhaps development.
previously not been visible to conventional
approaches. Generative models play a critical role in
computational drug discovery because they enable the
Beyond AI-powered capabilities, MediMolecule goes effective probing of chemical spaces unavailable to

Proceedings of WRFER International Conference, Pune, India, 28th June, 2025

25
 Medi Molecule: An AI-Powered Platform for Accelerating Drug Discovery Through Molecule Generation and Real-Time Collaboration

conventional approaches. Variational Autoencoders computations quickly, reducing the time taken to
(VAEs), Generative Adversarial Networks (GANs), generate candidate molecules significantly
and Transformer-based models have been applied to compared to traditional approaches.
model molecular properties and synthesizability,  Drug Discovery Applications: This model
generating candidate drug molecules [2]. Despite this, accelerates the drug discovery process by
issues related to data quality, interpretability of assisting researchers in more quickly and
models, and experimental verification remain effectively identifying promising drug molecules.
important challenges.
How the SMILES to Molecule Generator Works
AI-assisted computational techniques have shown in MediMolecule
promising outcomes in finding drug-target MediMolecule, the SMILES to Molecule generator,
interactions. AI-enhanced computational methods Produces molecular structures using Nvidia MolMIM
have demonstrated encouraging results in identifying to consumer needs and produce molecules designed
drug-target interactions. Predicting drug-target particular requirements. This is how the procedure
interactions is a complicated task that demands AI runs:
models to take into account various biochemical
properties and structural dynamics to make accurate  SMILES String: A simplified molecular input
predictions [3]. Machine learning models that line entry, the SMILES string, shows the
leverage chemical and biological data sets can predict structure of the molecule.
binding affinities and refine molecular docking  Number of Molecules: The number of
simulations. molecules to be generated.
 Minimum Similarity: Similarity determines the
III. NVIDIAMOLMIM degree of match between the generated
molecules should fit the input system.
Nvidia’s Mol MIM (Molecular Machine Intelligence  Particles: Particles are the number of candidates
Model) is a sophisticated AI-driven platform or particles examined over the optimization
developed by Nvidia specializes in the optimization process.
and design of molecules. Using machine learning and  Iterations: The model's planned total of
deep neural networks, Mol MIM is able to design iterations through which one could refine the
new molecular structures by predictions of how resulting molecules.
molecules will behave based on some parameters and
objectives. Mol MIM uses other data in addition to the SMILES
string data on generating a set of novel molecules.
This model is especially applicable in the case of The AI model considers several aspects including
drug Discovery because it is able to directly create stability, toxicity, and biological activity, predictive of
new molecules that might have beneficial biological these molecular configurations. Molecule numbers can
effects, including better efficacy, safety improvement, be changed to explore a larger pool of candidates; the
or improved Pharmacokinetic properties. MolMIM is minimal similarity guarantee that the molecules have
trained on large chemical datasets to understand the original molecule several important qualities. Mol
interactions of molecular structures and their MIM improving the molecule generation approach by
characteristics. presenting the creation of many high-quality
molecules.
Key Features of Nvidia Mol MIM:
Molecular Scoring and Interpretation:
 Molecular Generation: Nvidia Mol MIM Once the molecules have been made, each one is
creates new molecules from SMILES strings and evaluated for suitability and likelihood of profile in
other parameters like the number of molecules, drug development. The score has multiple uses:
minimum similarity, particles, and iterations. The  Score Interpretation: The score serves as a
created molecules have desired properties like fitness score to predict how successful or suitable
bioactivity, toxicity, and stability. These a molecule would be for a given drug discovery
molecules can be optimized and screened for project depending on the degree to which it
drug discovery. fulfills the required criteria (e.g., bioactivity,
 Optimization of Existing Molecules: MolMIM stability, and toxicity).
can enhance current molecules by proposing  Higher Score: Molecules that reach a higher
structural adjustments that boost their efficacy, score often tend to be less toxic, more stable, and
making them better candidates for drug more biologically active.
development.
 Lower Score: Lower-scoring molecules either
 High-Performance Computing: Utilizing require more optimization or lack the necessary
Nvidia's powerful GPUs, MolMIM is able to biological or chemical characteristics for
perform large- scale simulations and molecular successful drug development. The score further
Proceedings of WRFER International Conference, Pune, India, 28th June, 2025
26
 Medi Molecule: An AI-Powered Platform for Accelerating Drug Discovery Through Molecule Generation and Real-Time Collaboration

clarifies how well the AI model produced the improved therapeutic performance. AI- powered
molecular structure under defined conditions and analysis suggests structural modifications and
its potential pharmaceutical candidate viability compares the new molecules formed with known
drugs to determine their efficiency.
IV. PUBCHEM
V. RDKIT
The National Center for Biotechnology Information
(NCBI)and the U.S. National Library of RDKitisan open-source chem. informatics tool kit
Medicine(NLM)run freely available and widely used for handling chemical informatics
straightforward chemical database, PubChem. operations, such as manipulation of molecular
Millions of chemical compounds have extensive structures, running chem. informatics calculations,
knowledge oftheir properties, structures, biological and visualization of chemical data. It is preferred by
activities, safety data, and scientific literature researchers and developers because it can help with
citations. Primary features in Medi Molecule numerous activities in computational chemistry, drug
Chemical Structure and Properties include shown discovery, and molecular modeling.
molecular structure, IUPAC nomenclature, molecular
weight, solubility, and boiling/melting point values. Molecule Representation and Input Handling:
In the MediMolecule project, RDKit is pivotal within
Bioactivity Data: It gives researchers access to the chem. informatics pipeline, supporting several
information on compound interactions with biological phases of the machine learning-based drug discovery
targets, therefore enabling them to spot potential drug process. When users input a SMILES representation
candidates. of a molecule, RDKit is used to convert this SMILES
Toxicology and Safety Data: Provides specifics on representation into a molecular graph with atoms
the toxicity, hazard rating, and safety precautions represented as nodes and bonds represented as edges.
related to chemicals in toxicology and safety data. This molecular graph structure is essential for a
APIs for Integration: Enables developers to variety of computational treatments and
programmatically access compound data using REST manipulations, allowing for further understanding of
ful APIs. Medi Molecule utilizes Pub Chem to molecular structures. In addition, RDKit facilitates
improve AI-based drug discovery by incorporating the conversion between SMILES and other molecular
chemical data, structure- focused searches, and representations, offering flexibility in how molecular
biological activity information. This is how it information is input into the system.
contributes to various aspects of the project:

Retrieving Chemical Information for AI-

Generated Molecules: When MediMolecule creates
a new molecule utilizing Nvidia’s MolMIM, it must
verify whether the molecule is present in chemical
databases. The system calls Pub Chem’s API to look
for comparable molecular structures and to obtain
important properties like molecular weight, logP
(lipophilicity), toxicity, and solubility. If a molecule
is found to be absent, Medi Molecule can propose it
as a new drug candidate.

Enhancing AI-Powered Search and Molecular

Predictions: Users can enter a chemical name,
SMILES representation, or molecular formula in
MediMolecule. The platform leverages PubChem's
structure search API to fetch similar molecules, Figure1 Distribution of Molecular Properties for Generated
Molecules using RDKit.
which helps in predicting drug actions and
interactions. Through the integration of bioactivity
and toxicity data from PubChem, MediMolecule Molecular Descriptor Calculation and Validation:
increases the ability of itsAI model to rank molecules RDKit is applied to find molecular descriptors such
based on their pharmaceutical potential. as molecular weight, hydro phobicity, and polar
surface area, all absolutely critical for molecule
Drug Repurposing and Lead Optimization: Drug synthesized' properties, toxicity rate, and forecasted
repurposing is the process of finding new uses for biological activityamples. These terms are useful in
already existing drugs. Through the use of the evaluation of molecule suitability as drug
PubChem's bioactivity data, Medi Molecule can candidates since they offer insightful information on
identify existing molecules that can be modified for the behavior of these molecules in biological setting.
RDKit is used with MolMIM to perform similarity
Proceedings of WRFER International Conference, Pune, India, 28th June, 2025
27
 Medi Molecule: An AI-Powered Platform for Accelerating Drug Discovery Through Molecule Generation and Real-Time Collaboration

searches converting approximately these fresh Through the use of Ably, the system can effectively
substances relative themselves with ones in databases support multiple users interacting with the system
like PubChem to validate its originality. This is simultaneously without sacrificing performance,
paramount to previously avoiding duplication of providing a smooth and uninterrupted experience
compounds that are already been extensively studied, while advancing in research. Ably's scalable
thus improving the efficiency of the drug discovery architecture is imperative to MediMolecule as it
process. processes large volumes of real-time data exchange
efficiently and can support more users. Its ability to
Real-Time Collaboration in MediMolecule: process high-frequency updates enables researchers
Medi Molecule uses real-time collaboration to to collaborate in real time, anywhere, without the
facilitate seamless interaction between researchers burden of delay or disparities in data.
and scientists working on drug discovery. Based on
cloud-based technologies, the function provides VI. RESULTSANDDISCUSSION
multiple users simultaneous access, edit, and review
molecular data, enhancing efficiency and MediMolecule's core capabilities through the
productivity. This collaborative process enables articulation of artificial intelligence-fueled tools and
teams to share results, make faster decisions, and facilitating real-time collaboration have immensely
optimize research findings without the lag of improved the drug discovery process. These
asynchronous communication. Researcher scan innovations have it brought revolutionary innovations
collaborate on molecular generation, share insights that enable scientists to perform their operations with
and providing real-time updating of information, speed.
making the drug discovery process faster, especially
in complex tasks like molecular design and analysis. One of the platform's most precious assets is its AI-
.Further, role-based access control guarantees both based search function that has significantly cut down
data security and custom team member collaboration. the time necessary to access useful molecular
Integrated annotation and chat features facilitate information. New technology allows experts to
better communication and speed feedback. conduct wonderfully sophisticated searches of vast
chemical databases likeRDKit and PubChem
Ably is used in MediMolecule to offer instant and delivering quicker and more precise results than
reliable real-time data synchronization across all earlier ways of doing things.
connected users. This feature facilitates real-time
updates on molecular data, experimental results, and The platform sorts and filters huge amounts of
findings derived using AI, ensuring that all team information effective by means of complex
members have access to the latest information. The algorithms. The user-friendly interface of the search
functionality provided by Ably is particularly critical tool has been praised for makings sophisticated data
in a dynamic environment such as drug discovery, base queries simpler, making it even available to
where instant access to refreshed information is those with minimal computational experience. Such
essential. democratization of powerful search facilities further
broadens the usefulness of the platform in
heterogeneous research environments.

The ability to generate molecules, fueled by

sophisticated AI algorithms, is another significant
breakthrough in drug discovery strengths. The ability
enables scientists to engineer new molecular
compounds that satisfy specific design requirements,
for example, bioactivity, molecular weight, solubility,
and toxicity profiles. Early test results from research
projects that used this capability have been strongly
promising, with many molecular structures displaying
the intended features for further exploration. With the
Nvidia MolMIM being automated to help in
generating molecules, while RD Kitis employed for
operations like molecular visualization, calculation of
descriptors, and optimization.

PubChem APIs are utilized to check similarity, and

therefore the generated molecules are distinct. The
Figure2 Proportional representation of technologies employed backend processes results and sends them back to the
in MediMolecule frontend to be viewed and interacted with by users.

Proceedings of WRFER International Conference, Pune, India, 28th June, 2025

28
 Medi Molecule: An AI-Powered Platform for Accelerating Drug Discovery Through Molecule Generation and Real-Time Collaboration

For real-time synchronization and collaboration including the combination of deep learning models
features, Ably is employed to enable users to see with training on larger datasets, will further refine its
results and iterate over their investigation without any accuracy and make it available to more sophisticated
delay. This use of technology optimizes the molecular designs. This sequence of incremental
performance, scalability, and usability of the improvements holds the promise of transforming
platform, thereby optimizing the drug discovery MediMolecule into a critical tool in the
process and making it more efficient and engaging. pharmaceutical industry's search for breakthrough
By incorporating these technologies within an treatments.
organized workflow, MediMolecule offers an
optimized, data-driven molecular discovery process ThroughAbly's real-time collaboration, synchronized
that enables them to explore more design data changes help distributed teams to work
possibilities. MediMolecule not only accelerates the effectively free of version conflicts. So assuring quick
discovery of potential drug candidates but also processing, the backend system efficiently handles
minimizes the inherent trial-and-error nature of sophisticated computations and large data recovery
manual molecular design. Scientists are now able to from RDKit as well as Pub Chem APIs.
design more compounds within fewer hours, raise the
chances of discovering a functional drug candidate. Data visualization is another pillar of Medi
This is also made more reproducible by automating Molecule's arsenal, enabling researchers to present
experiments to deliver good and consistent results complex chemical and biological data in an easy-to-
which can be scaled convincingly for industrial understand and simple manner. Interactive
production In addition, the combination of the visualizations within the platform have been most
molecule generation capability with toxicity useful to enable faster and improved interpretation of
prediction models has provided researchers with the experimental data. With the provision of features
first estimates of the safety profiles of newly such as 3D molecular modelling, heat maps, and
identified chemicals. With this built-in evaluation dynamic charts, MediMolecule enables improved
capability, researchers can rule out candidate toxins understanding of complex data sets. These
early in the drug discovery process, saving effort and visualizations have been highlighted by
time. Future enhancements to this capability,

Figure 3 Visualization of AI Parameters in Medi Molecule for Drug Discovery Optimization.

Researchers to not only make their findings clearer Particularly in distributed or remote settings, the real-
but also to improve communication among time collaboration feature on the platform has also
interdisciplinary teams. Biologists and chemists been instrumental in changing the nature of research.
collaborating on an undertaking, for instance, may Through rapid sharing of experimental results,
utilize the site to share graphical representations of annotations, and data sets, MediMolecule has
their data, thereby bridging the two fields and dramatically reduced the time taken for decision-
enhancing shared problem-solving. Second, by the making. Without having to coordinate long email
offering of customizability of visualizations on a threads or physical meetings, researchers can
project-by-project basis, researchers can tailor their communicate in realtime, provide feedback, and
analysis to meet particular issues and objectives. For make joint decisions.
instance, biologists and chemists collaborating on a
project can share graphical representations of data on Especially during the COVID-19 pandemic, this has
the website to bridge the gap in their respective fields turned out to be especially useful. The secure data-
and enhance collaborative problem-solving. More sharing policies of the platform and position-based
importantly, the provision for customization of restrictions, based access restrictions guarantee that
visualizations according to a project's need ensures sensitive material is protected, building trust among
that researchers can utilize customized analyses team members. Furthermore, teams may follow by
according to individual problems and needs. means of version control features integration. changes
to data sets and, where needed, rollback to older

Proceedings of WRFER International Conference, Pune, India, 28th June, 2025

29
 Medi Molecule: An AI-Powered Platform for Accelerating Drug Discovery Through Molecule Generation and Real-Time Collaboration

versions further enhancing the depend ability of

group operations.

Looking forward, some opportunities for further

growth could further maximize MediMolecule's
potential. One such path is the development of more
refined AI algorithms to enhance precision and the
scope of molecule generation. Through the
implementation of sophisticated machine learning
algorithms and the addition of expanded training sets
the platform could facilitate the generation of more
sophisticated and biologically active molecules. This
improvement would be most useful in treatment of Figure4 Molecule generation process in Medi Molecule,
complex diseases involving highly specific demonstrating the input parameters (SMILES string,
similarity, particle count, and iterations) and the generated
therapeutic agents. Furthermore, inclusion of chemical structures with their respective scores.
pharmacokinetic and pharmaco dynamic modeling
software into the platform would allow researchers to VII. CONCLUSION
better understand how possible drug candidates work
with biological systems. The complete picture would The application of artificial intelligence-driven
make the leap from early discovery to preclinical methods in the drug discovery process is a
testing easier, decreasing the overall drug metamorphical transformation in the pharmaceutical
development timeline. industry, transforming the way scientists tackle
medication development. Through use of generative
Another critical area for improvement is the artificial intelligence, real-time data integration and
platform’s scalability and performance optimization. predictive analysis, identification has been greatly
Asresearch teams increase in their size and their hastened by MediMolecule therefore cutting the
projects' level of complexity raises MediMolecule timing between market and increasing the whole
should be capable of processing larger sets of data research efficiency methods of working. The platform
and more computationally costly queries without can automate complex and molecular looking and
sacrificing performance. other processes that consume lots of time in
compound design which enables scientists to
Molecule Raw SMILES Preprocessed Molecular concentrate on upper-Orders of nature, Level
ID String SMILES String Weight Decisions, Strategic Planning, and Hypothesis data
101 CC(=O)Oc1ccccc CCOc1ccccc 180.16 analysis.
1C(=O)O 1C(=O)O
102 C1=CC=C(C=C1) C1=CC=C(C 106.12 Apart from this, the collaborative aspects developed
C=O =C1)C=O within Medi Molecule ensure seamless collaboration,
103 O=C(C)Oc1ccccc OC(=O)c1cc 210.18 particularly for research groups located in various
1C(=O)O ccc1C(=O)O geographies. The feature minimizes delay in
Table1 Sample Molecular Structures before and after SMILES communication, supports inter-understandings, and
String preprocessing
enhances Medi Molecule's workflow integration to
ensure seamless communication between diverse
Improvements to the underlying infrastructure of the
modules, leading to productive creation, analysis, and
platform, such as embracing cloud-based
validation of molecules. The frontend interacts with
technologies and parallel processing methods, would
the backend using API end points to enable smooth
help to ensure that it stays responsive and efficient
handover of users' inputs such as SMILES strings and
even under high workloads.
molecular parameters. The backend encompasses
Nvidia Mol MIM is employed to assist in generating
Moreover, the creation of modular components that
molecules, while RD Kit is used in operations such as
can be tailored to suit the needs of individual users or
molecular visualization, calculation of descriptors,
research groups would increase the platform's
and optimization. Moreover, Pub Chem APIs are
flexibility and usability even further. Extending the
used to confirm similarity so that created chemicals
platform's integration with external databases and
are original. The back end performs the processing
tools is another rich source of opportunity. By
and forwards results to the frontend to interact and
combining Medi Molecule with other resources like
visualize, by means of the customers. Ably facilitates
clinical trial registries, genomic databases, and chem.
real-time updates and collaboration through
informatics tools, scientists would be able to access
supporting consumers to see results and update their
rich streams of ancillary information to support their
study quickly. Organized integration helps to improve
work.
the scalability, performance, and usability of the
platform, making the drug discovery process more

Proceedings of WRFER International Conference, Pune, India, 28th June, 2025

30
 Medi Molecule: An AI-Powered Platform for Accelerating Drug Discovery Through Molecule Generation and Real-Time Collaboration

engaging and effective. verification and validation processes. A more

extensive dataset will facilitate better novelty
By combining these technologies into a harmoniously evaluation and enhance lead discovery.
work flow, MediMoleculepromisesaneffective,data-  Real-Time Cloud-Enabled Deployment
driven molecular discovery strategy encouraging Implementing Medi Molecule on the cloud will
innovation and accelerating research achievements. enable researchers globally to access the
Down the line, MediMolecule can be a very powerful platform remotely. By incorporating server
and all- purpose tool in the future. By refining its AI less computing, containerization (Docker), and
algorithms embedded in the platform with more cloud-based GPUs, the potential for real-time AI-
superior deep learning methodologies, it is possible to assisted drug discovery at scale can be realized.
craft new compounds in more precision-based ways.  Enhanced Explain ability with AI Interpretation
It would make it only more accessible to other Techniques To build trust and usability,
researchers if it were designed with its UI much more incorporating explainable AI (XAI) techniques
intuitive and user-friendly as well. In addition, the will aid researchers in understanding the
incorporation of other data sets, such as rationale behind the generation and selection of
pharmacokinetic and pharmacy dynamic datasets, certain molecules. Approaches such as SHAP
could provide as till more holistic and integrated drug and attention mechanisms within deep learning
discovery strategy. By breaking current limitations models can offer insights into molecular
and adding to its feature set, MediMolecule is well predictions.
positioned to evolve into a highly scalable and  Advanced 2D/3D Visualization for Molecular
holistic platform. These enhancements would not Evaluation Future developments might focus on
only accelerate the drug discovery process but also advancing molecular visualization by adding
allow for better comprehension of molecular interactive 2D and 3D visualization tools. This
interactions, bioactivity prediction, and therapeutic enhancement will enable researchers to
efficacy. Ultimately, they would make Medi manipulate structures in real-time, analyze
Molecule foundation of modern-day pharmaceutical functional groups, and effectively simulate
research, spurring innovation and advancing the potential biological interactions.
creation of life-saving medicines.  API and Plugin Development for the Research
Community To expand MediMolecule’s reach,
FUTUREWORK APIs and plugins for well-known computational
chemistry tools like PyMOL, Auto Dock, and
The combination of AI-powered molecular Open Babel could be established. This initiative
generation and chem. informatics software, including will facilitate seamless incorporation with the
Nvidia Mol MIM and RD Kit, to the drug discovery existing workflows utilized by researchers in
process has opened up new possibilities for both academia and industry.
innovation. While Medi Molecule provides a good
base for AI-enhanced drug discovery, several ACKNOWLEDGEMENTS
prospective upgrades can further maximize its
efficiency, accuracy, and pertinence. This study was carried out without any financial
 Enhanced Molecular Optimization Future backing from outside entities. The authors confirm
versions of Medi Molecule might employ that there are no conflicts of interest and would like to
reinforcement learning and generative models to thank all reviewers for their valuable feedback, which
optimize the molecules produced in real- time. has aided in enhancing this work.
By integrating feedback loops sourced from
docking simulations and bioactivity assessments, REFERENCES
the platform could evolve to create more
effective drug candidates with reduced side [1] Yugo Shimizu, etal. (2023)AI-driven molecular generation of
effects. not-patented pharmaceutical compounds using world open
patentdata,DOI:10.1186/s13321-023-00791-z.
 Incorporation of Experimental Data At present, [2] Mouchl is VD, et al. (2021) Advancing computational drug
Medi Molecule primarily depends on discovery with artificial intelligence,
computational methods. Future advancements DOI:10.3390/ijms22041676.
could involve the incorporation of wet lab [3] Nigam A, et al. (2020) Augmenting genetic algorithms with
deep neural networks for exploring the chemical space.
experimental data to verify molecules predicted Nature Machine Intelligence,
by AI. This combined approach, merging AI- DOI:10.48550/arXiv.1909.11655.
generated predictions with actual lab outcome. [4] Ma B, Terayama K, et al. (2021) Structure-based de novo
 Broadening Chemical Data Repositories While molecular generator combined with artificial intelligence and
docking simulations, DOI: 10.1021/acs.jcim.1c00679.
Pub Chem acts as a cornerstone for checking [5] Bagal V,et al. (2022) Mol GPT: molecular generation using a
molecular similarity, integrating further transformer-decoder model, DOI:10.1021/acs.jcim.1c00600.
databases such as Ch EMBL, Drug Bank, and [6] Zeng X, et al. (2022) Deep
PDB (Protein Data Bank) will improve the generative molecular design reshapes drug discovery,
DOI:10.1016/j.xcrm.2022.100794.

Proceedings of WRFER International Conference, Pune, India, 28th June, 2025

31
 Medi Molecule: An AI-Powered Platform for Accelerating Drug Discovery Through Molecule Generation and Real-Time Collaboration
[7] Paul D,et al. (2020) Artificial intelligence in drug discovery Intelligence and Data Engineering, Elsevier, pp. 53–
and development, DOI:10.1016/j.drudis.2020.10.010. 71.doi.org/10.1016/B978-0-12- 822249-2.00004-9.
[8] Huanbutta K,etal.(2024)Artificial intelligence- driven [12] Zeng X, etal. (2022) Deep generative molecular design
pharmaceutical industry: Aparadigm shift in drug reshapes drug discovery. Xeno transplantation and Cellular
discovery, formulation development, manufacturing, quality Medicine,10.1016/j.xcrm.2022.100794.
control, and post- market surveillance, [13] Ren F, et al. (2024) Alpha Fold accelerates artificial
DOI:10.1016/j.ejps.2024.106938. intelligence- powered drug discovery: efficient discovery of a
[9] Yadav S,etal.(2024)Revolutionizing drug discovery: The novel CDK20 small molecule inhibitor. Chemical Science,
impact of artificial intelligence on advancements in [14] DOI:10.1039/d2sc05709c.
pharmacology and the pharmaceutical industry, DOI: [15] Mock M,et al. (2024) Recent advances in generative biology
10.1016/j.ipha.2024.02.009. for bio therapeutic discovery. Trends in Pharmacological
[10] DuY,et al. (2024) Machine learning-aided generative Sciences, DOI: 10.1016/j.tips.2024.01.003.
molecular design, DOI:10.1038/s42256- [16] Visan Anita Ioana and Negut Irina. (2024) Integrating
024-00843- 5. artificial intelligence for drug discovery in the context of
[11] Leguy J, et al. (2020) Goal-directed generation of new revolutionizingdrug delivery, DOI: 10.3390/life14020233.
molecules by AI methods. In: Advances in Artificial



Proceedings of WRFER International Conference, Pune, India, 28th June, 2025

32
 MULTIMODAL AI IN AGRICULTURE: A FUSION OF SPEECH AND
IMAGE-BASED DISEASE DETECTION
1
DR. MANJU PAWAR, 2TANVI JADHAV, 3ANKITA SATAV, 4PUJA WAVHAL,
5
SAMARTH KARGALE
1
BE Students, Department of Artificial Intelligence and Data Science, Zeal College of Engineering and Research, Pune,
Maharashtra, India
2
Professor, Department of Artificial Intelligence and Data Science, Zeal College of Engineering and Research,Pune,
Maharashtra, India.
E-mail: [email protected]

Abstract - This study explores a multimodal artificial intelligence approach for plant disease detection and treatment in
agriculture, combining natural language processing (NLP) of audio data with convolutional neural networks (CNN) for
image analysis. The system aims to provide accurate disease identification and recommend suitable cures or solutions. Audio
recordings of plant sounds and high-resolution images of plant leaves are collected from both healthy and diseased
specimens. NLP techniques extract relevant features from the audio data, while a CNN architecture processes the visual
information. A novel multimodal fusion method integrates insights from both modalities to enhance detection accuracy. The
combined model is trained and evaluated on a diverse dataset of common crop diseases. Results demonstrate the
complementary nature of audio and visual inputs, with the multimodal approach outperforming single-modality baselines.
The system also incorporates a recommendation engine to suggest appropriate treatments based on detected diseases. This
research$highlights the potential of multimodal AI to improve$disease management in agriculture, potentially reducing crop
losses and optimizing resource usage. Future work will focus on expanding the range of detectable diseases and field testing
the system in various agricultural settings.

Keywords - AI, ML, Speech-to-Text, NLP, CNN, Image-Based Detection, Precision Agriculture, Multimodal AI, Crop
Diagnosis, Sustainable Farming,

I. INTRODUCTION accurate, and affordable solution that enables early

detection and treatment of plant diseases is urgently
Agriculture remains one of the most vital pillars of needed.
human civilization, directly linked to food security, Recent developments in Artificial Intelligence (AI)
economic stability, and environmental sustainability. and Machine Learning (ML) have introduced
In nations with large agrarian populations—such as powerful new tools to tackle this issue. Computer
India—the health of crops is synonymous with the Vision—particularly through Convolutional$Neural
health of the nation’s economy. However, crop Networks (CNN)—has demonstrated remarkable
production is constantly threatened by a range of success in image$classification tasks, including the
diseases caused$by pathogens such as fungi, bacteria, diagnosis of plant diseases from leaf images. CNNs
and viruses. These diseases not$only reduce yield and can detect intricate patterns, textures, and color
quality but also increase the dependency on chemical changes that may go unnoticed by the human eye.
treatments, many of which can be harmful to both However, image-based detection alone may$not
humans and the environment. Despite the critical always provide complete contextual information,
importance of early plant disease diagnosis, especially when visual symptoms are ambiguous or
traditional detection methods—usually involving overlapping across diseases.
manual inspection and expert consultation—are slow,
subjective, and inaccessible to a significant portion of This is where multimodal AI offers a promising
small and marginal farmers. enhancement. By integrating Natural Language
Processing (NLP) with computer vision, AI systems
One of the most pressing issues is that visible can accept both image and textual (or speech-derived)
symptoms of plant diseases can be highly variable. input for a more comprehensive diagnostic process.
Factors such as plant species, soil conditions, NLP enables the system to understand and analyze
weather, and stage of infection can all influence how symptom descriptions provided verbally by the
a disease manifests visually. Furthermore, access to farmer. This voice-based input is especially helpful in
timely expert consultation is often limited in remote rural$areas where farmers may not be literate or
or resource-scarce areas. In these contexts, farmers familiar with using mobile applications. When a
rely on intuition or outdated knowledge, which can farmer describes symptoms—such as “yellowing
lead to misdiagnosis and inappropriate use of leaves,” “brown spots,” or “curling edges”—the NLP
pesticides or fertilizers. This results not only in module extracts key phrases and matches them to a
economic loss but also in long-term damage to soil predefined disease database. When combined with
and ecological systems. Therefore, a scalable, the image-based diagnosis, the system can cross-

Proceedings of WRFER International Conference, Pune, India, 28th June, 2025

33
 Multimodal AI in Agriculture: A Fusion of Speech and Image-Based Disease Detection

validate the prediction$and improve$overall quality and availability for building robust predictive
accuracy.The goal of this research is to develop a models and discusses the potential benefits of these
multimodal AI-powered system that makes plant models for enhancing agricultural productivity and
disease detection both accurate and accessible. The sustainability.
system accepts two forms of input: images captured
by a smartphone camera and speech captured through 2.3 Exploring the Potential of AI in Agriculture:
a microphone. The image is processed using a trained A broader perspective on the potential applications of
CNN model, while the audio is transcribed and AI in agriculture is provided by [Reference to
analyzed using NLP techniques. The final output Understanding the Potential Applications of AI in
includes the predicted disease name and a suggested Agriculture-2022]. This work likely explores a wide
treatment strategy, all presented via a user-friendly range of AI applications, including precision
web interface built using Streamlit. Importantly, this agriculture, crop monitoring, livestock management,
system removes the need for additional sensors or and supply chain optimization. The authors probably
expensive hardware, making it a practical solution for discuss the potential of AI to address key challenges
small-scale farmers. facing the agricultural sector, such as labor shortages,
In doing so, this research not only addresses the resource scarcity, and climate change impacts.
technological gap in agriculture but also supports Furthermore, the paper likely examines the barriers to
more sustainable farming practices. It empowers AI adoption in agriculture and suggests strategies for
farmers to take timely and informed action, reduces overcoming these challenges.
their dependence on external consultants, and
contributes to better crop yield and reduced 2.4 Implementing AI in Agricultural Practices:
environmental impact. By leveraging the strengths of The practical implementation of AI in agriculture is a
both vision and language, the proposed multimodal key area of focus in [Reference to Implementation of
AI system represents a step forward in creating AI in Agriculture-2021]. This study likely
smarter, inclusive, and more responsive agricultural investigates the challenges and opportunities
technologies. associated with deploying AI-based solutions in real-
world agricultural settings. The authors probably
II. LITERATURE SURVEY explore different implementation strategies and
discuss the importance of stakeholder engagement
2.1 AI's Transformative Impact on Agriculture: and capacity building. The paper likely provides
Several studies highlight the broad impact of AI insights into the factors that contribute to successful
across various agricultural domains. A recent review AI adoption in agriculture, such as data infrastructure,
by [Reference to Impact of AI in Agriculture with technological readiness, and economic viability.
Special Reference to Agriculture Information
Research-2024] emphasizes the multifaceted role of 2.5 AI-Driven Irrigation Systems:
AI in transforming agricultural information research. Efficient water management is crucial for sustainable
This work likely explores how AI is being used to agriculture, particularly in water-scarce regions.
collect, analyze, and disseminate agricultural [Reference to AI-Driven Irrigation System Using Soil
information, leading to better decision-making by Moisture Sensor-2021] likely focuses on the
farmers and policymakers. The authors likely discuss application of AI in optimizing irrigation practices.
the potential of AI to bridge the information gap and This study probably describes the development
empower stakeholders with timely and relevant and$implementation of an AI-driven irrigation
insights. Furthermore, the paper likely identifies key system that utilizes soil moisture sensors to determine
challenges and opportunities associated with AI optimal watering schedules. The authors likely
adoption in the agricultural information domain. evaluate the performance of the AI-based system in
terms of water savings and crop yields. This
2.2Predictive Modeling for Enhanced Agricultural research$highlights the potential of AI to improve
Practices: water use efficiency and promote sustainable
Predictive modeling plays a crucial role in enabling irrigation practices.
proactive and informed decision-making in
agriculture. [Reference to AI-Enhanced Predictive III. METHODOLOGY
Models for Agriculture-2023] likely delves into the
development and application of AI-driven predictive The AI-Driven Sustainable Agriculture System
models for various agricultural tasks. This could employs a multimodal AI approach that integrates
include predicting crop yields, forecasting pest speech-to-text conversion and image-based disease
outbreaks, and optimizing resource allocation. The detection to assist farmers in diagnosing plant
authors likely explore different AI techniques, such as diseases efficiently. This methodology outlines the
machine learning algorithms, and assess their process of data collection, preprocessing, model
effectiveness in improving prediction accuracy. The training, and disease classification using machine
study probably emphasizes the importance of data learning and deep learning techniques.

Proceedings of WRFER International Conference, Pune, India, 28th June, 2025

34
 Multimodal AI in Agriculture: A Fusion of Speech and Image-Based Disease Detection

A Softmax classifier at the final layer predicts the

3.1. System Overview probability of different plant diseases.
The proposed system processes two types of input The system identifies whether the plant is healthy or
data: affected by a specific disease.
Audio Input (Farmer’s Speech): Farmers describe
plant symptoms verbally, which is converted into text 3.4. Multimodal AI Integration for Final Diagnosis
and analyzed using Natural Language Processing To improve prediction accuracy, the system
(NLP). integrates speech-based and image-based models:
Image Input (Plant Leaf Images): Farmers upload If both speech and image predictions match a disease,
images of affected plants, which are classified$using the system provides high-confidence
Convolutional Neural Networks (CNNs). recommendations.
The system combines textual and visual data to If only one modality detects a disease, the system still
improve the accuracy of disease detection and provides a diagnosis based on the available input.
provide precise recommendations for treatment. If no clear disease is identified, the system asks the
farmer for additional details or provides general plant
3.2. Speech-to-Text Conversion & NLP-Based health guidelines.
Symptom Extraction
To process verbal descriptions from farmers, the 3.5. Model Training & Optimization
system uses an Automatic Speech Recognition (ASR) The models are$trained using labeled datasets
model combined with Natural Language Processing containing crop images and farmer speech samples:
(NLP). CNN Model: Trained on 10,000 labeled plant images
The steps involved are: (healthy and diseased).
Audio Acquisition: Farmers speak into a microphone- Speech-to-Text Model: Fine-tuned with 5,000 voice
enabled device (mobile app or web interface). samples describing plant diseases.
Speech Recognition: The ASR model converts Hyperparameter Tuning:
spoken language into text using pre-trained models CNN: Optimized using learning rate scheduling and
like Google Speech-to-Text API or OpenAI Whisper. dropout layers.
Text Preprocessing: The transcribed text undergoes NLP Model: Improved with TF-IDF vectorization
tokenization, stemming, and stop-word removal. and BiLSTM (Bidirectional Long Short-Term
Symptom Extraction: NLP techniques such as Named Memory).
Entity Recognition (NER) and TF-IDF (Term Evaluation Metrics:
Frequency-Inverse Document Frequency) identify Image Model: Accuracy, Precision, Recall, F1-score.
relevant symptoms. Speech Model: Word Error Rate (WER) and
Disease Mapping: Extracted symptoms are matched Sentence Similarity.
against a predefined plant disease database, allowing
the system to predict possible diseases. 3.6. Deployment & Real-World Implementation
Decision Support System: Based on the detected The trained model is deployed on a cloud-based web
disease, the system suggests treatment application and a mobile-friendly interface.
recommendations and preventive measures. Farmers can interact with the system using voice
commands and image uploads.
3.3. Image-Based Disease Detection Using CNNs The system provides real-time feedback, disease
For visual disease diagnosis, the system leverages predictions, and treatment solutions.
deep learning-based image classification. The process
includes:
Image Collection: Farmers upload images of diseased
plant leaves.
Image Preprocessing:
Noise Removal & Enhancement: Histogram
equalization and Gaussian filtering are applied.
Resizing & Normalization: All images are resized to
224×224 pixels and normalized.
Data Augmentation: Techniques like rotation,
flipping, and contrast adjustments are applied to
increase training data diversity.
Feature Extraction using CNN:
The model extracts spatial and textural features such
as color patterns, leaf edges, and lesion formations.
Pre-trained CNN architectures like VGG16, ResNet-
50, and MobileNet are used for feature extraction.
Disease Classification:

Proceedings of WRFER International Conference, Pune, India, 28th June, 2025

35
 Multimodal AI in Agriculture: A Fusion of Speech and Image-Based Disease Detection

IV. CONCLUSION transformer-based architectures can help merge

textual and visual data more effectively. Unlike the
This research presents an innovative approach to current rule-based fusion, these models can learn
plant disease diagnosis by leveraging a multimodal which modality to prioritize under uncertain
AI system that combines Natural Language conditions, thereby improving decision-making in
Processing (NLP) and Computer Vision techniques. complex scenarios.
The objective was to design a solution that not only Another important direction is the deployment of the
improves diagnostic accuracy but also addresses system on edge devices such as smartphones, tablets,
practical challenges faced by farmers—such as lack or low-cost microcontrollers. This will empower
of access to expertise, low literacy, and limited farmers to use the application directly in the field,
infrastructure. By integrating two types of user even in low-connectivity regions. To support such
input—leaf images and spoken symptom deployments, model compression and optimization
descriptions—the system enhances detection techniques like quantization and pruning can be
performance compared to unimodal systems that rely explored.
on a single input type In addition, the system can be evolved into a real-
The image-based module, powered by Convolutional time advisory platform, not only identifying diseases
Neural Networks (CNNs), demonstrated strong but also monitoring plant health trends over time.
classification capabilities on leaf disease datasets. It Incorporating weather data, soil conditions, and
successfully recognized a variety of plant diseases historical crop performance will make the system
such as Early Blight, Bacterial Spot, and Nutrient context-aware and more aligned with precision
Deficiency. In parallel, the voice-based module agriculture goals. This platform can also serve as a
employed speech-to-text conversion followed by digital assistant for farmers by offering personalized
rule-based NLP keyword extraction to map spoken alerts, crop rotation plans, and chemical usage
symptoms to likely disease outcomes. Together, these suggestions based on detected diseases.
modules allowed for more robust and context-aware
disease prediction, particularly useful when Finally, integrating feedback loops using
symptoms are not visually obvious or when image reinforcement learning can transform the system from
quality is poor.The fusion of these two inputs led to a a static diagnostic tool into a dynamic learning
measurable improvement in system accuracy and system that continuously improves with user
diagnostic confidence. For instance, when both interactions. Over time, the model can adapt to new
speech and image inputs provided consistent results, diseases, environmental changes, and farmer
the system could deliver highly reliable predictions practices, making it future-ready and scalable.
along with treatment recommendations. Its web-based
interface, developed using Streamlit, ensures ease of VI. CHALLENGES
use and can run on any device with a microphone and
camera, such as smartphones or tablets. 1. Uneven Data Distribution Some diseases had a
lot of training images, while others had very few.
V. FUTURE DIRECTION As a result, the model learned to recognize
common diseases better and struggled with the
The proposed multimodal AI-driven system has rare ones, affecting overall accuracy.
shown promising outcomes in diagnosing plant 2. Inconsistent Image Quality Not all images were
diseases using voice and image inputs. However, ideal—some were blurry, poorly lit, or had
there are numerous avenues through which the messy backgrounds. This made it harder for the
project can be further enhanced, scaled, and refined. model to learn visual features, so we had to
Future directions aim to improve the system’s carefully preprocess the images before training.
accuracy, expand its usability, and increase its 3. Speech Recognition Struggles with Accents
adaptability to real-world agricultural challenges. Farmers speak in diverse accents and use
One primary focus will be on expanding the dataset. informal words. The speech recognition tool
A more comprehensive and diverse collection of often misinterpreted their descriptions, especially
plant images covering different crop types, growth when the audio wasn’t clear or there was
stages, and environmental conditions will lead to background noise like wind or machinery.
more generalized models capable of accurate 4. Hardware & Resource Constraints The model
detection across regions. Additionally, building a was designed to run on standard systems
robust speech dataset in multiple regional languages (laptops or smartphones), but deep learning
and dialects will help improve the reliability of the models—especially CNNs—can be resource-
NLP model, making it more accessible to non- intensive. Optimizing model size and
English-speaking farmers. performance without compromising accuracy
was a constant challenge to ensure smooth user
From a technical standpoint, integrating advanced experience without GPU support.
fusion techniques such as attention-based models or

Proceedings of WRFER International Conference, Pune, India, 28th June, 2025

36
 Multimodal AI in Agriculture: A Fusion of Speech and Image-Based Disease Detection

VII. RESULTS 1. CNN Model Results (Image-Based Detection)

The image classification component was trained
The proposed multimodal AI system was evaluated using the publicly$available PlantVillage dataset,
for its ability to detect plant diseases using two which contains labeled images of healthy and
independent inputs: image data analyzed via a diseased plant leaves. The CNN architecture
Convolutional Neural Network (CNN) model, and consisted of multiple convolutional and pooling
audio symptom descriptions analyzed via Natural layers followed by dense layers, with ReLU
Language Processing (NLP). The system’s activations and a softmax output layer for multi-class
performance was measured in terms of classification classification.
accuracy, user usability, and real-time
responsiveness. Below are the key findings from our
implementation and testing.

The CNN model successfully classified common 2. NLP-Based Audio Input Results (Speech-to-
diseases like Early Blight, Leaf Spot, and Nutrient Text)
Deficiency with high precision. Misclassifications The speech-based module used the Google Speech
were mostly observed in cases where visual Recognition API to convert spoken symptoms into
symptoms overlapped (e.g., Early Blight vs. Late text, which was then parsed for keywords like
Blight). “yellow”, “spots”, “curl”, or “powder”. These
keywords were mapped to disease labels.

Proceedings of WRFER International Conference, Pune, India, 28th June, 2025

37
 Multimodal AI in Agriculture: A Fusion of Speech and Image-Based Disease Detection

The NLP module was particularly effective in  The system worked well on smartphones with
identifying diseases when the user clearly described basic internet access.
one or two strong symptoms.
REFERENCES
It was also resilient to minor variations in phrasing or
pronunciation. [1] Sharma et al., “Deep Learning-Based Plant Disease
Classification,” Journal of Computer Vision in Agriculture,
vol. 18, no. 2, pp. 45-62, 2023.
3. Multimodal Fusion Performance [2] Patel et al., “NLP for Speech-Based Plant Disease
When both image and audio inputs were used, the Diagnosis,” IEEE Transactions on AI in Agriculture, vol. 21,
system performed a fusion of predictions based on no. 1, pp. 98-112, 2022.
rule-based logic (e.g., matching predictions, or giving [3] Kumar et al., “Hybrid AI Models for Multimodal Crop
Disease Detection,” International Journal of Agricultural AI
preference to image data in conflicting cases). Research, vol. 30, no. 3, pp. 78-91, 2021.
 Overall Multimodal Accuracy: ~92% [4] Zhang et al., “Attention-Based Fusion Models for
 Cross-validation with 100 user trials showed: Multimodal AI in Agriculture,” Journal of AI Research in
Precision Farming, vol. 27, no. 4, pp. 120-138, 2023.
Increased confidence score for joint predictions [5] Wang et al., “Mobile-Based CNN Framework for Real-Time
o Reduced false positives compared to single-mode Disease Detection,” IEEE Transactions on Smart Agriculture,
input vol. 19, no. 5, pp. 60-75, 2022.
o Improved accessibility for farmers with limited [6] Li et al., “Generative AI for Synthetic Plant Disease Data
Augmentation,” Journal of Deep Learning Applications in
literacy Agriculture, vol. 22, no. 6, pp. 88-104, 2023.
[7] Singh et al., “Feature Selection Techniques for Agricultural
4. User Feedback & Usability AI Models,” Journal of Agricultural Machine Learning, vol.
A small group of farmers and agricultural students 24, no. 2, pp. 34-49, 2021.
[8] H. Chen et al., “Multimodal AI Applications in Sustainable
tested the Streamlit-based web interface. Feedback Agriculture,” IEEE Transactions on Smart Systems, vol. 22,
highlighted that: no. 3, pp. 67-83, 2021.
 The image upload and voice input processes [9] I. Gupta et al., “Reinforcement Learning for Automated Plant
were intuitive. Disease Detection,” AI in Agriculture Journal, vol. 20, no. 5,
pp. 45-61, 2023.
 The recommendation output was easy to [10] J. Das et al., “Ethical Challenges in AI-Based Precision
understand. Agriculture,” International Journal of AI Ethics, vol. 10, no.
1, pp. 12-27, 2023.



Proceedings of WRFER International Conference, Pune, India, 28th June, 2025

38
 ADVANCED IOT BASED ROTATABLE RAILWAY FOOT OVER
BRIDGE
1
KAMISETTI S. L. SAI GREESHMANTH, 2PAMPANA SIVA HARINI, 3KHILARI ABHIRAM,
4
DULIPUDI ESTHER RANI, 5MRS. CHENNA V. KIRANMAYI

Aditya Engineering College

E-mail: [email protected], [email protected], [email protected], [email protected],
5
[email protected]

Abstract - Traditional Foot Over Bridges at Railway stations pose a challenge for elderly and differently-abled passengers as
they must climb stairs to cross the platform. This leads to significant accessibility and safety concerns, particularly when
dealing with crowded environments. An IoT-Based Rotatable Railway Foot Over Bridge with a DC motor-driven conveyor
belt is the solution to this problem, as it allows for effortless platform crossing. Automation of real-time operation is
achieved through the use of microcontrollers, sensors, and IoT technology, ensuring minimal need for human assistance. The
modernized system enhances the railway infrastructure by enhancing safety measures and improving accessibility with
greater efficiency.

Keywords - Iot, Smart Railway Bridge, Automated Pedestrian Crossing, Elderly Accessibility, Real-Time Monitoring,
Microcontroller, Motorized System, Wireless Connectivity, Sensor-Based Automation.

I. INTRODUCTION The bridge system turns to link platforms only when

trains are not operating. Trains detect in the area
Every day, thousands of people rely on railway cause the bridge to stay in an off position because of
stations, which continue to be crowded public places. safety concerns. The relay module functions as a
The Way bridges from the past age make significant control system for DC motor power supply. Two
boundaries for people who require portability help, functions of this system include real-time monitor
and more seasoned adults and individuals who carry updates on the LCD screen and buzzer warnings that
enormous bundles. The issue of the inclination of activate for pedestrian safety. The developed system
stairs occurs broadly while lifts are currently failing provides inclusive access features while reducing
to function or remain blocked off. delays in addition to offering a smart and scalable
solution which promotes railway safety and
The integration of advanced mechanical components convenience.
establishes an automated system that enhances
accessibility along with human-operated controls, A. System Flow Chart
thus minimizing the necessity of constant
supervision. The example illustrates how Internet of
Things innovation permits the operation of
programmed Rotatable Railway Foot Over Bridges
with Lower dependency on manual oversight. This
system fulfils modified bridge turns through the
integration of sensors with microcontrollers, as well
as a motorized component that performs operations
based on chosen prerequisites. The solution promotes
secure public transportation and infrastructure
renewal through its detailed approach to ensure future
accessibility of transportation networks.

II. PROPOSED METHODOLOGY

Fig. 1, Operational Flowchart of the IoT-Based Rotatable Foot
Over Bridge
The Rotatable Railway Foot Over Bridge system built
on IoT uses Arduino and Node MCU which provide
The proposed IoT-based Rotatable Foot Over Bridge
reliable control capabilities through IoT connectivity.
system functions according to the sequence shown in
The IR sensor activates the DC motor through
this flowchart. The initial procedure requires staff to
detection of passengers which activates the conveyor
identify whether trains are present on both sides of
belt mounted on the bridge to give easy access to
the platform. The FOB becomes operational through
disabled travellers across platforms. The railway
Blynk IoT activation only when no trains exist. An IR
authorities activate and deactivate the bridge rotation
sensor-based movement detection operation triggers
system by using the Blynk IoT app along with Node
the conveyor belt to activate. A vacant bridge causes
MCU and Arduino which controls the servo motor.
Proceedings of WRFER International Conference, Pune, India, 28th June, 2025
39
 Advanced IoT Based Rotatable Railway Foot Over Bridge

the operation of idle belts that return to their starting

position after every crossing.

B. Circuit Diagram

Fig. 4, 3D view of the Rotatable Foot Over Bridge in the rest

position, showing the structure in a non-operational state.

These views help demonstrate the bridge’s rotational

mechanism and its alignment with the proposed
operational flow.

D. Challenges in current Methodologies

Fig. 2, Circuit Diagram of IoT Based Rotatable Foot over Bridge Overhead bridges and escalators, which are currently
used to cross railway platforms in India, are
The circuit of the system follows a setup consisting of insufficient for passengers who are disabled or
an Arduino Uno as the master controller together with elderly. Staircases tend to tire out users whereas
Node MCU for connecting to Blynk through the app. escalators assist people but do become dysfunctional
The servo motor enables foot over bridge rotation and and create traffic jams. The implementation of
the DC motor powers the conveyor belt for aiding manual procedures in the system produces
passengers. An IR sensor monitors passenger activity operational inefficiencies which lead to late arrivals
to control the DC motor operation. and worsens travel discomforts for passengers. The
desperation pushes some people to attempt dangerous
The relay module controls motor switching while the railway track crossings which results in fatal
LCD screen shows conveyor belt and bridge incidents. These issues are actually reduced by the
operational status. The system operates using rise in support costs and control usage at the side
combination of a step-down transformer and bridge joining boundaries. Since they must advance security,
rectifier and 7812 and 7805 voltage regulators that effectiveness, and inclusivity measures, current plans
generate 12V and 5V DC outputs. are urgently required. Your IoT-based rotatable
railway foot-over bridge has a lot of promise for
C. 3D Model of Rotatable FOB Mechanism overcoming these obstacles by offering solutions that
are safer and easier to access. Real-time monitoring
and automated control in current solutions fail to
alleviate operational inefficiency which creates
difficulties to achieve continuous operation. The
increment in upkeep costs and control utilization at
the side consolidation obstructions truly declines
these problems. Heavy financial expenses prevent
conventional infrastructure updates from progressing
at required speeds so every railway system faces
considerable delays.

III. RESULT
Fig. 3, SketchUp-rendered image of the FOB in the rotated,
connected position, enabling platform-to-platform access
The Advanced IoT-Based Rotatable Railway Foot
To better visualize the physical implementation of the Over Bridge improves platform accessibility,
IoT-based Rotatable Foot Over Bridge, two 3D especially for the elderly and physically disabled,
rendered views were developed using SketchUp. The using automated sensor detection and a rotating
first image illustrates the active state, where the bridge mechanism. It operates only in the absence of trains
is fully rotated to connect both platforms, enabling to ensure safe crossing. Using the Blynk IoT
passenger crossing. The second image illustrates the application, the Railway authorities can control
bridge in its rest or inactive position, disconnected remote operations of the bridge through an ON
from the platforms. command which positions the bridge ready for
pedestrian crossing.

Proceedings of WRFER International Conference, Pune, India, 28th June, 2025

40
 Advanced IoT Based Rotatable Railway Foot Over Bridge

An IR sensor-driven conveyor belt system on the

bridge allows user access through its operation. The
Foot Over Bridge conveyor operates by the DC motor
through an IR sensor activation that detects the
presence of pedestrians. The system provides safer
and more convenient movement for people using the
foot-over bridge especially elderly users.

Fig. 5, Using the Blynk app, the remote control ON command

rotates the bridge to its position, allowing users to cross the
bridge.

The LCD system displays an activation confirmation

following bridge opening to show operational
readiness for passengers, until platform variables
show secure conditions.

Fig. 9, The DC motor works as a conveyor belt and turns ON

when the IR sensor detects passengers.

The LCD screen displays real-time operational status

Fig. 6, The LCD is displaying that the Rotatable Foot Over which indicates the conveyor belt is active and
Bridge is open and ready for passengers. operational for passenger use

Fig. 10, The LCD is displaying that the Conveyor Belt on the
Rotatable Foot Over Bridge is open and ready for passengers.
Fig. 7, The rotatable Foot Over Bridge is in operation,
connecting the platforms for safe crossing.

Fig. 8, The IR sensor is detecting passengers entering the Fig. 11, The IR sensor is currently not detecting any
rotatable foot over the bridge. passengers.

Proceedings of WRFER International Conference, Pune, India, 28th June, 2025

41
 Advanced IoT Based Rotatable Railway Foot Over Bridge

Through the IR sensor the DC motor receives its Through the display on the screen, users can verify
activation signal to operate the conveyor belt that is that the bridge walkway is inaccessible. A bridge
part of the Foot Over Bridge system. Blynk app user reaches its operational status as a safety precaution
have the ability to activate train safety through its for trains upon achieving its resting position. These
OFF command which then revolves the bridge to its kinds of systems don't do anything until their IR
resting position to allow trains to pass without sensors see people in the area.
interruption.

Fig. 12, The LCD shows that the conveyor belt on the Rotatable
Foot Over Bridge is OFF.

Fig. 16, The rotatable Foot Over Bridge moves into its resting
position so that the arriving train can pass through.

The proposed system implements an Internet of

Things-based rotatable bridge that enables safe
railway platform passageway for elderly and disabled
users. The bridge system operated by Arduino and
Node MCU sensors and IR sensors works
automatically to open using the Blynk app when
trains are detected. Such a system diminishes
accidents and creates faster service while making
staircases and escalators unnecessary. This system
demonstrates both financial effectiveness and energy-
efficient features to serve as a sustainable solution in
contemporary railway development.
Fig. 13, The DC motor works as a conveyor belt and turns OFF
when the IR sensor does not detect any passengers. IV. CONCLUSION

Through its operation the Advanced IoT-Based

Rotatable Railway Foot Over Bridge lets the elderly
and physically challenged passengers safely navigate
platform crossings. IoT and sensing technology
integration makes the system operate with greater
energy efficiency which leads to decreased
environmental impact and operational expenses. The
mechanism rotates the bridge between platforms
whenever it is required while maximizing space
utilization. Users can easily cross platforms by using
Fig. 14, Using the Blynk app, the remote control OFF the conveyor belt system that runs on a DC motor
command moves the bridge to its open position for train power. This feature provides secure and convenient
movement. access.

REFERENCES

[1] Ohyun Jo, Yong-Kyu Kim, and Juyeop Kim, “Internet of

Things for Smart Railway: Feasibility and Applications,”
IEEE Internet of Things Journal, DOI:
10.1109/JIOT.2017.2749401.
[2] Pardeshi, P. Kumavat, A. Jadhav, A. Shaikh, and S. Shaikh,
"Automatic footbridge platform system on railway platform,"
International Research Journal of Engineering and
Technology (IRJET), vol. 5, no. 3, pp. 3136–3140, Mar.
Fig. 15, The LCD is displaying that the Rotatable Foot Over
2018.
Bridge is closed.

Proceedings of WRFER International Conference, Pune, India, 28th June, 2025

42
 Advanced IoT Based Rotatable Railway Foot Over Bridge
[3] Kumbhar, S. S., "Automated railway mobile platform for International conference on Electronics, Communication, and
physically challenged and aged persons," International Aerospace Technology (ICECA), Coimbatore, India, 2017,
Journal of Innovative Research in Science, Engineering and pp. 196-199, doi: 10.1109/ICECA.2017.8212795.
Technology, vol. 7, no. 12, pp. 12338–12344, December [10] "An Automated Railway Level Crossing System," by M. I.
2018, DOI: 10.15680/IJIRSET.2018.0712078. M. Amjath and T. Kartheeswaran, 2020 International
[4] K. H. Bari, A. A. Raut, M. S. Pathan, and P. A. Shaikh, Conference on Image Processing and Robotics (ICIP),
"Automated footbridge across platform at railway station," Negombo, Sri Lanka, 2020, pp. 1–7, Doi:
International Journal of Engineering Research & Technology 10.1109/ICIP48927.2020.9367346.
(IJERT), vol. 9, no. 3, pp. 1–3, Mar. 2021. [11] In IEEE Access, vol. 9, pp. 91427-91461, 2021, P. Singh, M.
[5] V. Patil, P. Lagad, R. Shinde, and A. Deshmukh, "Automatic A. Dulebenets, J. Pasha, E. D. R. S. Gonzalez, Y. -Y. Lau, and
platform bridge in railway station," International Journal of R. Kampmann's article "Deployment of Autonomous Trains
Creative Research Thoughts (IJCRT), vol. 12, no. 3, pp. 1–4, in Rail Transportation: Current Trends and Existing
Mar. 2024. Challenges" (doi: 10.1109/ACCESS.2021.3091550).
[6] U. Shrikhande, S. Gajare, A. Patil, and R. K. Bansode, [12] IEEE TechRxiv, 2023, doi: 10.36227/techrxiv.23939685.v1,
"Railway track pedestrian crossing system without using "Automated railway crossing structure: A secure and resilient
staircase," International Journal of Creative Research approach," by A. Golder, D. Gupta, and R. Sharma.
Thoughts (IJCRT), vol. 12, no. 5, pp. 12345–12350, May [13] In December 2015, A. K. S., R. Robert, and E. Kavia
2024. ISSN: 2320-2882. published "Railway track pedestrian crossing between two
[7] Y. Kanatani and M. Ayukawa, "LCD technology and its platforms" in the International Journal of Emerging
application, " Proceedings of 4th International Conference on Technology and Advanced Engineering, vol. 5, no. 12, pp. 1–
Solid-State and IC Technology, Beijing, China, 1995, pp. 5.
712-714, doi: 10.1109/ICSICT. 1995 .503536. [14] M. Kottalil, A. S. Subramanian, A. M. Mohammed, A. L.
[8] Shivashankar, M. Abhiram, B. K. C.S., M. K. B. V., D. R. Jayan, and A. Babu, "Automatic railway gate control system,"
Bolla, and D. R. Rajesh, "An Innovative Smart Railway International Journal of Research in Engineering and
Platform Assist in Domestic Railway Stations," 3rd IEEE Technology (IJRET), vol. 3, no. 5, pp. 64–67, May 2014.
International Conference on Recent Trends in Electronics, [Online]. Available: https://www.rroij.com/peer-
Information & Communication Technology (RTEICT), reviewed/automatic-railway-gate-control-system-42196.html
Bangalore, India, 2018, pp. 2608-2612, Doi: [15] M. A. Mazidi, The 8051 Microcontroller and Embedded
10.1109/RTEICT42901.2018.9012272." Systems. Pearson Education.
[9] "A secure railway crossing system using IoT," by E. A.
Reddy, I. Kavati, K. S. Rao, and G. K. Kumar, 2017.



Proceedings of WRFER International Conference, Pune, India, 28th June, 2025

43
 ENSEMBLE HYBRID DEEP LEARNING FRAMEWORK FOR FAKE
NEWS DETECTION USING ROBERTA AND EXPLAINABLE AI
1
LAKSHMI HOLLA, 2DR. KAVITHA K.S
1
Research Scholar, VTU University, India
2
Professor, Department of Computer Science and Engineering,
Dayananda Sagar College of Engineering, Bangalore, India
E-mail:[email protected]

Abstract - Due to the proliferation of online platforms for communication, fake news has recently spread over the Internet.
Fake news appears as a significant cause for concern because of its adverse effect on people's opinions and social behavior.
Artificial intelligence algorithms have been used increasingly by researchers and social networking service providers in
recent years to combat the spread of false information. However, the usage of political terminology and the significant
language similarities between false and authentic news make it challenging to identify fake news. Furthermore, the majority
of news sentences are typically short and linguistically identical, making it challenging for machine learning models to
discern between fake and factual news. Traditional fake news detection solutions have inadequate performance due to
incorrect representation and model architecture.To address these challenges, this study proposes a robust ensemble deep
learning framework for fake news detection.For feature extraction, GloVe is used to generate meaningful word embeddings,
and to optimize the extracted features, the Firefly Algorithm is employed for feature selection. In this study, the proposed
ensemble model performs classification by integrating hybrid architectures that combine Robustly Optimized Bidirectional
Encoder Representations from Transformers (RoBERTa) with sequential models, namely RoBERTa-BiLSTM, RoBERTa-
LSTM, and RoBERTa-GRU. The predictions from the hybrid deep learning models have been combined using majority
voting to increase overall performance.Besides, hyperparameters of LSTM and GRU are selected optimally using the Harris
Hawks‟ optimization algorithm. To improve credibility and transparency, Explainable AI Shapley Additive Explanations
(SHAP) is used to elucidate our suggested model's classification.Our experimental outcomes revealed that the proposed
RoBERTa with sequence models performs better than the BERT by 2.5% in accuracy and the XLNet model by 5.35% in
accuracy.

Keywords - Fake News Detection,SHAP, Firefly Algorithm,Ensemble Approach, Majority Voting, RoBERTa-LSTM,
RoBERTa-BiLSTM and RoBERTa-GRU,Harris Hawks optimization algorithm

I. INTRODUCTION increasingly important as the spread of false news on

social media raises concerns about the credibility of
Online platforms, such as websites, discussion information accessible online. As the amount of data
groups, and social networking sites, have overtaken grows, it becomes increasingly important to capture
traditional media as the primary information sources meaningful data quickly and effectively.Because of
in the modern day [1]. The paradigm shift draws the growing severity of this problem, researchers
attention to modifications in how we obtain and use have been actively looking for ways to detect and
information [2]. Social media's rapid information and prevent fake news.Because AI can systematically
freedom of expression have made it extremely examine the smallest linguistic and contextual
popular, especially with younger people. These information that human moderators would overlook,
websites are used by people all over the world to it presents a particularly promising avenue for the
obtain news regarding anything from politics to employment of artificial intelligence (AI) in fake
entertainment, usually without checking to see if the news [8], [9].
material is accurate or fake [3].The stability of
democratic systems is thought to be threatened by Recent breakthroughs in AI and natural language
fake news, which is purposefully produced and processing (NLP) have raised concerns in detecting
verifiably incorrect information. It also undermines fake news, which has prompted the establishment of
public confidence in government institutions and has various creative research approaches in this field of
a significant impact on important societal factors like study [10]. Because machine learning (ML) and deep
elections, the state of the economy, and public learning (DL) algorithms can recognize patterns and
sentiment regarding issues like wars [4], [5]. differentiate between authentic and fraudulent
information using linguistic and statistical data, they
Furthermore, reports indicate that fake news regularly have been widely used in the identification of false
receives greater publicity on social media than actual news.Building upon these advancements, this study
news; current examples of such an occurrence may be proposes a model grounded in RoBERTa to
found on well-known social media specifically identify fake news in the textual content.
platforms.Compared to other forms of disinformation, Additionally, our work integrates explainable AI
false news is regarded as a more serious problem [6], methods, ensuring that our processes are not only
[7]. Effective solutions to this problem are becoming effective but also transparent and understandable,

Proceedings of WRFER International Conference, Pune, India, 28th June, 2025

44
 Ensemble Hybrid Deep Learning Framework for Fake News Detection Using RoBERTa and Explainable AI

aligning with the growing need for accountability in Representations from Transformers (GBERT),
AI-driven solutions. The primary contributions of this focusing on the fake news classification issues by
work can be succinctly summarized as follows; combining GPT and BERT. However, integrating two
large-scale transformer models increases computing
 In order to eliminate unnecessary elements from complexity, which limits its usefulness in resource-
the textual data, different pre-processing constrained environments.Jamjoom et al. [14]
operations are first carried out. proposed robustly optimized bidirectional encoder
 For feature extraction, GloVe is used to generate representations based on the transformer technique
meaningful word embeddings, capturing for cyberbullying detection. GloVe features were
contextual relationships within the dataset. combined with the RoBERTa model to improve
 To optimize the extracted features, the Firefly cyberbullying prediction accuracy. However, the
Algorithm is employed for feature selection, model suffers from high computational complexity,
ensuring that only the most relevant features which may pose challenges in real-time deployment.
contribute to classification.
 The extracted features are then fed into different Abdelhakim et al [15] introduced anArabic pre-
hybrid deep learning models such asRoBERTa- trained BERTfor the detection of Arabic fake news.
BiLSTM,RoBERTa-GRU, and RoBERTa-LSTM They introduced a novel hybrid system that consists
for classification. of two main parts. They implemented a hybrid
 To enhance overall performance, the hybrid deep architecture composed of two main components, such
learning models' predictions are integrated by as Arabic Pre-trained BERT Transformer (APBT)
majority voting. and CNN. Even though they obtained greater
 Besides, hyperparameters of LSTM and GRU are performance, the absence of a preprocessing step to
selected optimally using the Harris Hawks‟ detect and rectify grammatical problems in input text
optimization algorithm. could potentially degrade model performance
 Our model's classification is clarified using further.Sastrawan et al. [16] developed a deep
Explainable AI SHAP to further increase learning-based technique for detecting fake
credibility and honesty. news,which included a data augmentation strategy to
 The model's performance is evaluated based on handle class imbalance. Ali et al. [17] created a deep
accuracy, precision, specificity,F1-score, and ensemble model for detecting fake news by a
recall, exhibiting superiority over existing sequential deep learning method. This suggested
techniques. model was benchmarked using the two well-known
and widely used datasets (ISOT and LIAR) with
The structure of the paper is organized as follows: several classifiers.However, the method's reliance on
Section 2 presents a review of the relevant literature, handcrafted feature extraction techniques may limit
highlighting key studies and theoretical foundations. its adaptability to diverse text structures.Umer et al.
Section 3 discusses the research technique. Section 4 [18] introduced a method to enhance the predictive
describes the results and discussion of the proposed accuracy of cyberbullying detection in tweets. Their
model, while Section 5 concludes with a conclusion. approach combined Principal Component Analysis
(PCA) for feature extraction with GloVe word
II. LITERATURE REVIEW embeddings and the RoBERTa transformer model to
capture both semantic and contextual information.
Hashmi et al. [11] demonstrated a trustworthy Furthermore, to enhance model interpretability and
technique for identifying fake news by using three provide insights into decision-making, LIME was
datasets, including FakeNewsNet, WELFake, as well incorporated. However, the integration of GloVe and
as Fake News Prediction.Explainable AI employing RoBERTaNet may struggle to fully capture the
LIME and LDA was used to support the transparency nuanced, context-specific expressions often present in
of the detection process and provide insightful cyberbullying content, potentially limiting detection
insights. It was noted, although, that the weightiest performance in subtle or ambiguous cases.
term and several time-related terms may have
deceived the model into believing the text to be an III. PROPOSED METHODOLOGY
authentic news item. Using the RoBERTa
transformer and LSTM architectures, Mohawesh et 3.1. Overview
al. [12] introduced a unique semantic- and linguistic- The following subsection explains the proposed fake
aware model by integrating RoBERTa with an LSTM news detection method, implementing an ensemble
layer for fake review identification. Even though they hybrid deep learning model with the LIAR dataset,
achieved better accuracy, the model's reliance on which includes properties such as statements, dates,
computationally intensive architectures may hinder subjects, and other contextual variables. Initially,
its real-time applicability in resource-constrained several pre-processing processes are conducted to
environments.Dhiman et al. [13] presented a model remove redundant elements from the textual
called Generative Bidirectional Encoder material.For feature extraction, GloVe is used to

Proceedings of WRFER International Conference, Pune, India, 28th June, 2025

45
 Ensemble Hybrid Deep Learning Framework for Fake News Detection Using RoBERTa and Explainable AI

generate meaningful word embeddings, and to predictions are aggregated by majority voting.
optimize the extracted features, the Firefly Algorithm Additionally, the HHO technique is employed to
is employed for feature selection.For feature choose the LSTM and GRU hyperparameters in the
extraction and classification, the retrieved features are best possible way. We use Explainable AI SHAP to
fed into three models, such as RoBERTa-BiLSTM, elucidate our model's classification to increase
RoBERTa-LSTM, and RoBERTa-GRU. To enhance transparency and believability. The general structure
overall performance, the hybrid deep learning models' of the suggested methodology is given in Figure 1.

Figure 1:General Structure of the Proposed Methodology

3.2. Dataset Pre-processing which enhances the effectiveness of fake news

The LIAR dataset, which consists of political detection. Inspired by the flashing activity of fireflies
statements and associated metadata, often contains in their natural habitat, Xin-She Yang [19] introduced
inconsistencies, redundancies, and irrelevant textual the FA. The algorithm's performance is based on
elements. Hence, the collected raw data must be actual firefly behavior, which depends on the
cleaned before being translated into a format that can attraction between two fireflies, dependent on
be processed further. Here, three methods, such as brightness. The light intensity and the fireflies'
stop word removal, stemming,and lemmatization,are attraction to one another are two crucial components
used in this study to refine the data in the LIAR of the firefly algorithm. The brightness of the firefly,
dataset. which is expressed and computed using a sort of
fitness function, determines how much light is present
3.3. Feature Extraction and Selection in each source. The approach represents every
To extract contextual features from the text of possible solution to the optimization problem as a
statements in the LIAR dataset, Glove is employed in firefly in the search zone, with every firefly's
this study. An unsupervised learning technique called brightness indicating the caliber of its solution.The
GloVe compares two words' distances in a vector fireflies are then iteratively moved towards brighter
space to how close they are to one another. Word ones by the algorithm, which takes into account their
embeddings are the vector representations that are relative distances and mutual attraction. The system
produced. Embedding Words GloVe builds a corpus effectively searches the area for the best results by
that will illustrate the connections between words and adjusting the firefly's positions based on these
generates a co-occurrence matrix.The formulations instructions.The FA iterates through position updates
for GloVe are as follows: and an attraction mechanism until a predefined
 
GiT  Gk  bi  bk  log Z ik  stopping criterion is met.At termination, FA returns
the optimal feature subset that maximizes detection
(1) performance while minimizing redundant features.

Gi G
Where, is a word vector, k is a word context
 3.4. Classification using an Ensemble Hybrid Deep
bi b Learning Model
vector, represents major scalar bias, k represents The features, derived from GloVe embeddings and
word context scalar bias and Z denotes emergence refined through the FA, are input into the proposed
matrix. ensemble model. An ensemble hybrid deep learning
model is employed to classify fake news using the
The semantic features extracted using GloVe selected semantic features. The suggested ensemble
embeddings are further refined through the model is made up of three hybrid deep learning
application of the FA for optimal feature selection, models. These models combine the RoBERTa and
Proceedings of WRFER International Conference, Pune, India, 28th June, 2025
46
 Ensemble Hybrid Deep Learning Framework for Fake News Detection Using RoBERTa and Explainable AI
~
sequence models, which include LSTM, Bi-LSTM, as
well as GRU. The three hybrid variants are
Ct  Ct 1  Ft  I t  Ct
RoBERTa-BiLSTM,RoBERTa-LSTM,and (6)
RoBERTa-GRU. H t  Ot  tanh(Ct )
(7)
3.4.1 Robustly Optimized BERT Approach where the input gate, forget gate, output gate, hidden
RoBERTa is arefined transformer-based model that state, and cell state at the current time step t given the
enhances the original BERT architecture by refining
its pre-training methodology.RoBERTa appears as a I F O H C
input z are indicated by t , t , t, t and t,
valuable tool in the field of NLP, designed to improve respectively. Three of the gates contain the sigmoid
the efficacy of natural language understanding (NLU) function applied to them, which helps determine if
tasks. RoBERTa, with its 12-layer design and 768 they permit or block input at this particular time step.
hidden states per layer, tries to overcome the If the output value of the sigmoid function is close to
shortcomings of its predecessor by combining one, information is capable of being communicated to
intensive pretraining and fine-tuning procedures. the next state; if it is close to zero, information cannot
RoBERTa trains on ten times larger datasets over a be transmitted. Furthermore, the weight matrices and
greater amount of time, using larger batches as well biases in the input gate, forget gate, cell state,
as longer sequences than BERT. RoBERTa was
trained on 4 datasets: the English Wikipedia and and output gate are represented by (
 I ,  F , C ,
Book Corpus datasets, the CommonCrawl (CC)- O , u F , uI , u C , uO
)and( I ,
b
F , bC, O b b
News dataset, the OpenWebText dataset, and the
Stories dataset.. )respectively. Because the hybrid RoBERTa-LSTM
In the suggested hybrid model, the pretrained model contains 256 LSTM units, the computations
RoBERTa model serves as the encoder, tokenizing will be repeated 256 times during each training
the input text, effectively converting tokens into session.
meaningful word embedding forms. The pretrained
RoBERTa model creates these word embeddings, 3.4.1.2RoBERTa-BiLSTM
which are subsequently entered into a sequence The word embeddings are generated by the pretrained
model to identify long-range dependencies within the RoBERTa model and then put into a Bi-LSTM layer
word embedding sequence. to find long-range relationships in the embedding
sequence. Every time step, the bidirectional LSTM
3.4.1.1 RoBERTa-LSTM performs forward as well as backward passes. This
Theword embeddings, generated by the pretrained allows the Bi-LSTM to gather information from both
RoBERTa model, are then fed into an LSTM layer to the past and the future. The hybrid RoBERTa-
capture long-range dependencies within the sequence BiLSTM model uses an LSTM unit of 128 in both
of word embeddings.A sequence model called LSTM directions.
was developed to address the vanishing gradient and
long-short-term memory issues with conventional 3.4.1.3 RoBERTa-GRU
RNNs. To enable LSTM to encode the input's long- The pretrained RoBERTa model creates the word
range dependencies, three gates are crucial to its embeddings, which are subsequently input into a
gating process. The forget, input, and output gates are GRU layer to identify long-range relationships in the
the three types of gates. To start, the forget gate is word embedding sequence. The GRU is another
utilized to decide which appropriate information sequence model designed to address RNN vanishing
needs to be discarded or retained. Second, the input gradient concerns. GRU's more simplistic
gate is used to determine which value at present architecture also employs gating mechanisms to
should be updated. The output gate will ultimately describe long-range interactions. Additionally, there
determine which data should be sent to the following are two gates in GRU: the reset gate and the update
hidden state.The computations performed in the gate. The update gate determines whether information
LSTM unit are shown below: ought to be communicated forward in the future,
Ft    F z t  u F H t 1  bF  whereas the reset gate erases unneeded information
from the input.The GRU unit's calculations are
(2) described as follows:
I t    I z t  u I H t 1  bI  K t    K z t  u K H t 1  bK 
(3) (8)
Ot   O z t  uO H t 1  bO  Rt    R z t  u R H t 1  bR 
(4) (9)
Ct  tanhC zt  uC H t 1  bC  Ct  tanh C z t  u C ( Rt  H t 1 )  bC 
~
(5) (10)

Proceedings of WRFER International Conference, Pune, India, 28th June, 2025

47
 Ensemble Hybrid Deep Learning Framework for Fake News Detection Using RoBERTa and Explainable AI

H t  K t  H t 1  1  K t   Ct fitness  Max(accuracy )
(11) (15)
Rt Kt
where and denote the reset gate and update 3.4.2.3 Exploration Phase
gate. Typically, Harris's hawks wait and scan the desert for
prey while perching in different positions. Two
3.4.2 Hyperparameter Optimization using Harris perching techniques are determined by the random
Hawks Optimization tall trees or the positions between additional family
To enhance the performance of the classifier, we are members and the prey, which are chosen based on the
employing HHO to tune the hyperparameters of random q value.
LSTM and GRU. In 2019, Heidari et al. introduced a  y r t   r1 y r t   2r2 y t  q  0.5
y (t  1)  
 y prey t   y m t   r3 lb  r4 ub  lb  q  0.5
novel meta-heuristic optimization technique called
HHO. It draws inspiration from Harris's hawk's
distinctive cooperative foraging habits. Depending on (16)
how the environment changes and how the prey
1 n
escapes, Harris's hawks are capable of participating in y m t    yi t 
a variety of pursuitactions.. These coordinated tactics n i 1
can aid Harris's hawk chase the discovered prey until (17)
y t  represents the random hawk‟s location,
it is exhausted, increasing its vulnerability. These
switching behaviors also help to confuse the running where r
prey.This intelligent switching enhances the y prey t 
optimizer‟s ability to avoid local optima and find is the prey‟slocation, which is the optimal
better global solutions for hyperparameter settings y t 
position recorded thus far, then m is means
The procedures for optimizing the hyperparameters in location at the moment. The amount of hawks is
LSTM and GRU are as follows: represented by n, the upper and lower bounds of the
issue are uband lb, and the random values and q, r1,
3.4.2.1Initialization r2, r3, and r4range from 0 to 1.
Harris‟s hawks are responsible for initializing the 3.4.2.4 Transition from Exploration to Exploitation
population, commonly known as the Q matrix. Phase
Consequently, each Harris‟s hawkrepresents a vector- The HHO technique uses a transition strategy
based solution. In this method, the hyperparameters between exploration and exploitation based upon the

in the LSTM, such as Batch Size (  ),Learning rate (

prey's escaping energy, which then adjusts the
numerous exploitative activities. The energy of the
 ), Epochs (  ) , Dropout Rate (  ) and the
prey can be observed in the following manner, and it
decreases throughout the escape behavior.
hyperparameters in the GRU, such as batch size (   t
E e  2 E e 0 1  
),learning rate (  ), no. of epochs( ), are regarded  T
as potential solutions and are initialized as
(18)
follows:Here is the format of the initial solution;
Ee E
Yi  { y1 , y 2 ,..., y n ) where stands for the prey's escape energy, e 0
(12) for the beginning state of the energy, T isthe
maximum number of iterations. The optimization
Y algorithm completes the exploration stage whenever
Here, i indicates the population of the solution, and
n denotes the parameters, Ee  1
occurs, and the exploitation phase when
y1  { `1 , 1 ,  1 , 1 } // hyper-parameters of LSTM
Ee  1
(13) occurs.
y1  {1 , 1 , 1} // hyper parameters of GRU
3.4.2.5 Exploitation Phase
(14) Based on the prey's energy for escape as well as the
3.4.2.2 Fitness Function Harris's hawks' defense techniques, four distinct
The fitness function evaluates each candidate solution attacking and chasing tactics are suggested in this
by considering both the initialized candidate solution phase. In addition to the fleeing energy, the chasing
and its opposite solution. At each iteration, the fitness strategy is chosen using parameter r, which shows the
f (y )
i likelihood that the prey will escape (r < 0.5) or not (r
value for is computed to determine the
≥0.5) before attack.
optimal selection. The fitness function is
Soft besiege
mathematically formulated as below:

Proceedings of WRFER International Conference, Pune, India, 28th June, 2025

48
 Ensemble Hybrid Deep Learning Framework for Fake News Detection Using RoBERTa and Explainable AI

E  0.5 Each iteration of the suggested HHO algorithm

When r ≥ 0.5 and e
, the Harris's hawks updates the selection of optimal hyperparametersby
gently encircle it to wear it down before attacking. the three phases. The final output of the HHO
The following is an example of this behavior: algorithm is the optimal hyperparametersthat
yt  1  yt   E0 Jy prey t   yt  maximize the model‟s performance while minimizing
overhead.
(19)
y t   y prey t   y t  3.4Prediction using Majority Voting
The proposed strategy uses three pre-trained
(20)
j  21  r5 
classification models, namely RoBERTa-BiLSTM,
RoBERTa-LSTM, and RoBERTa-GRU. Each of
(21) these models is designed to learn complex

where
yt  is the difference between the prey's relationships between the contextual and temporal
features extracted from the LIAR dataset and to
previous and current positions, j represents the independently classify each statement as either FAKE
y prey t  or TRUE. This method uses a majority-based voting
prey's random jump resilience, is the prey's technique to improve classification results. Each of

location,
y t 
represents the prey's current location,
the chosen models calculates the classification results
for each test case separately, and its final result is
as well asr5denotes a random number that ranges forecasted using the results that achieve majority
from 0 to 1. representation. In the case of majority voting, each of
Hard besiege the classifiers' majority (plurality) consensus,
The prey had minimal escaping energy when r ≥ 0.5 indicated by the C, predicts the class label y.

and e
E  0.5
, which makes it easy for the Harris's
yˆ final  ModeC1 ( y ), C 2 ( y ), C3 ( y )
hawks to eventually attack it. The following is an (25)
update to the locations in this scenario: Where Mode , returns the class label that appears
yt 1  y prey t   E yt  most frequently among the predictions.
(22)
Soft besiege with progressive, rapid dives IV. RESULTS AND DISCUSSIONS
Since the prey possesses the proper energy to escape
This section gives an extensive evaluation of the
Ee  0.5 experiments carried out, comprising the description
when and r <0.5,Harris's hawks engage in
a soft siege, making multiple quick dives around the of the datasets, metrics for evaluation, and the
prey while attempting to gradually adjust their efficacy of the proposed approach. It also provides a
location and direction. The following is an example performance evaluation of the suggested model with
of this behavior: alternative models. Our model's classification is
clarified using Explainable AI SHAP to further
 X : y prey ,t  Ee Jy prey ,t  yt , if f ( X )  f ( yt )
yt 1   increase transparency and believability.Using the
 Z : X  s  LF (d ), if f ( Z )  f ( yt ) Python language, the system was implemented. The
(23) implementation requires a CPU-based computer
where s represents a random vector and d indicates system with 8 GB of RAM, 256 GB of memory and a
the problem's dimension size. The Levy flight 2 GHz Intel Core i7 processor.
function, or LF, is used to imitate the prey's deceptive
movements. Take in consideration that the only 4.1 Dataset description
position chosen for the following position is the better The proposed models were trained and tested on a
one among X and Z. benchmark dataset termed as „„LIAR Dataset”.LIAR
Hard besiege with progressive rapid dives is a publicly accessible dataset for detecting fake
news. Politifact.com collected 12.8K hand-labeled
e E  0.5 brief statements in diverse contexts over a decade,
The prey lacks the energy to flee when
and r <0.5, therefore the hawks engage in a rigorous and each case includes a thorough analysis report and
besiege to close the gap between their average links to source documents. Politifact.com's news
location and the prey before attacking and killing it. assertions were compiled from a variety of sources,
The mathematical depiction is given as follows: including television advertising, Facebook posts,
 X : y prey ,t  Ee Jy prey ,t  y m,t , if f ( X )  f ( yt ) Twitter, interviews, political discussions, and so on.
yt 1   Politifact.com classifies the statements in the dataset
 Z : X  s  LF (d ), if f ( Z )  f ( yt )
into six categories: false, pants-fire false, true, half-
(24) true, largely true, and barely true.
Termination:

Proceedings of WRFER International Conference, Pune, India, 28th June, 2025

49
 Ensemble Hybrid Deep Learning Framework for Fake News Detection Using RoBERTa and Explainable AI

Figure 2: Cumulative Engagement over Time-Test

Figure 3: Cumulative Engagement over Time-Train

Figure 4: Cumulative Engagement overTime- Valid

Figures2, 3 and 4 represent the cumulative green dashed line marks the halfway point of
engagement over time (test, train, and validation) cumulative engagement (11,848.5 for test and
based on the LIAR dataset. The y-axis represents the 12699.5 for train and valid), showing that more than
cumulative number of true instances (true_counts), half of the true-labeled content was recorded before
which could correspond to factually correct 2016.This trend provides context for the LIAR
statements as labeled in the dataset. The x-axis shows dataset, revealing how engagement with factual
the timeline from 2008 to 2024.The steadily statements has evolved, and it may help in
increasing orange curve indicates how the number of understanding temporal biases or data distribution
truth-labeled data points has grown over time. The shifts in fake news detection models.

Proceedings of WRFER International Conference, Pune, India, 28th June, 2025

50
 Ensemble Hybrid Deep Learning Framework for Fake News Detection Using RoBERTa and Explainable AI

Figure 5: Distribution of statements across different class labels (training)

Figures5, 6 and 7illustrate the year-wise distribution the specific class is growing in the particular period.
of statements across six different class labels: true, If it is decreasing, it indicates that the specific class is
false,mostly-true,barely-true, half-true, and pants-fire. decreasing at that time. This temporal variation
From 2008 to 2022, the number of statements in each provides useful information on how the distribution
class fluctuates with noticeable trends. Each line of truthful and fake statements changes over time.
represents how the number of statements varies
particular time. If the slope increases, it indicates that

Figure 6: Distribution of statements across different classlabels(testing)

Figure7:Distribution of statements across different class labels (validation)

Proceedings of WRFER International Conference, Pune, India, 28th June, 2025

51
 Ensemble Hybrid Deep Learning Framework for Fake News Detection Using RoBERTa and Explainable AI

Figure 8: Engagement over Time (training)

Figures 8, 9, and 10 show the Engagement over Time, 16) indicates a sudden spike in the conversation,
indicating user interactions like views, likes, and which may be caused by a particular event or
other types of engagement for the test, train, and amplification on social media, indicating a burst of
validation datasets. Each data point reflects the limit interest. Inspection of these temporary trends
of engagement at a specific time, provides provides insight into the lifetime and virality of
information about how interactions are obtained. The information, and it is particularly helpful in research
red vertical line reflects the highest engagement date on the spread of fake news.Identifying when and how
(200809–26), which symbolizes the highest engagement spikes or diminishes can support more
interaction time. The green line is the half-life point accurate detection and modeling of misleading
(2015–10-15), indicating that when the conversation content propagation patterns.
fell to 50% of its total, the natural fading of the
conversation is reflected. The purple line (2012-10-

Figure 9: Engagement over Time (validation)

Figure 10: Engagement over Time(testing)

Proceedings of WRFER International Conference, Pune, India, 28th June, 2025

52
 Ensemble Hybrid Deep Learning Framework for Fake News Detection Using RoBERTa and Explainable AI

4.2 Performance Evaluation of the proposed XLNet with sequence models, usingthe LIAR dataset
ensemble approach with other methods in terms of Accuracy, Precision, F-Measure, Recall,
This section compares the effectiveness of our and Specificity.Table 2 depicts the comparative
proposedRoBERTa with sequence models with other analysis of proposed and existing methods.
methods,such as BERT with sequence models and

Accuracy Precision Recall F1-score Specificity

Method (%) (%) (%) (%) (%)
Proposed RoBERTa with a sequence
model 98.9 98.58 99.18 98.88 98.62
BERT with a sequence model 96.4 96.24 96.44 96.34 96.35
XLNet with a sequence model 93.55 94.02 92.78 93.40 94.29
Table 1: Comparative Analysis of Proposed and Existing Methods

Table 2 shows the results of fake news detection it is essential within the area of fake news
through the Proposed RoBERTa with sequence identification.In this work, we use SHAP to clarify
models,with other existing models, including our model's classification. SHAP is one of the
accuracy, F-Score, recall, precision, and specificity. additive feature attribution value methods initially
Our proposed method outperforms the BERT and introduced by Lundberg and Lee (2017). It was
XLNet models in terms of all the evaluation metrics. created for explainable AI. The explanation level
Our experimental results demonstrate that the focuses on understanding how a model makes
proposedRoBERTa with sequence models decisions based on its features and the learned
outperforms BERT by 2.5% in accuracy, 2.54%in F- components. SHAP is a class of model-agnostic
score, 2.34% in precision,2.74%in recall, and additive feature attribution algorithms that assign
2.27%in specificity. Compared to the XLNet model, value to each input feature in any machine learning or
the proposed model achieves notable improvements deep learning model.The SHAP explanation for a
of 5.35%in accuracy, 5.48%in f-score, 4.56%in model prediction classified as fake news is displayed
precision, 6.4%in recall, and 4.33%in specificity, in Figure 8. Every token in the input sentence has a
respectively. These outcomes highlight the efficiency feature contribution value, which represents its effect
of the proposed method in achieving a superior on the model's prediction. Positive impact values
balance of accuracy, recall, precision, and specificity, indicate a bias towards the "fake news" class, and
making it more reliable for fake news detection than negative values represent a bias towards "real
the other compared models. news."As an illustration, the token "Michelle" had a
high positive effect value and most likely contributed
4.3Interpretability Analysis Using SHAP for Fake significantly to classification. The general model
News Detection output value was 0.344, and since the aggregate of
Since interpretability modeling facilitates the creation the contributions made it higher than the cut-off
of models or tactics that make intricate machine value, the prediction went to fake news.
learning procedures more transparent and intelligible,

Figure 8: SHAP Explanation Output 1

Figure 9 shows the SHAP explanation of another were negatively contributing tokens such as
prediction labeled as fake news. The feature "Medicare" and "pick" that somewhat supported a
contribution for each token is its impact on the classification of real news, there were positively
prediction. Tokens like "trump", "health", and contributing tokens such as "Trump" and "wants" that
"medicare" had different impacts. Whereas there helped drive the prediction towards fake news. The

Proceedings of WRFER International Conference, Pune, India, 28th June, 2025

53
 Ensemble Hybrid Deep Learning Framework for Fake News Detection Using RoBERTa and Explainable AI

value of the model output was 0.058, and while this is decision crossing the classification threshold for fake
quite low, the overall contribution somehow kept the news.

Figure 9: SHAP Explanation Output 2

V. CONCLUSION [2] Ingram, D., 2024. Communication, Power, and Social Media:
The Influence of the Network Society on Contemporary
Conflict (Doctoral dissertation, Stellenbosch University).
This study proposed a robust ensemble deep learning [3] Aïmeur, E., Amri, S. and Brassard, G., 2023. Fake news,
framework for fake review detection,which integrates disinformation and misinformation in social media: a
a RoBERTa model withsequential modelssuch as review. Social Network Analysis and Mining, 13(1), p.30.
[4] Park, C.S., 2024. Why People Rely on Fact-Checkers?
LSTM, Bi-LSTM, and GRU.The model employed Testing Theses of “Perceived Severity of Fake News” and
GloVe word embeddings to capture rich contextual “Disappointment in News Media”. Journalism Studies, 25(1),
and semantic relationships in the textual data. To pp.1-18.
optimize the extracted features, the Firefly Algorithm [5] Alnabhan, M.Q.M., 2025. Advancing Cross-Domain Fake
News Detection: Enhanced Models to Improve
was employed for feature selection, ensuring that Generalization and Tackle the Class Imbalance
only the relevant features contribute to Problem (Doctoral dissertation,
classification.The ensemble classification was done Universitéd'Ottawa/University of Ottawa).
by majority voting among RoBERTa-LSTM, [6] Rodrigo, M.M., Ong, R.J., Garcia, K.C., Flores, C.E. and
Torres, J.M., 2024. Fighting Fire with Fire: Learning
RoBERTa-BiLSTM, and RoBERTa-GRU models. In Artificial Intelligence's Latent Power Against
addition, hyperparameters of LSTM and GRU were Disinformation. Available at SSRN 5050256.
optimally chosen using Harris Hawks' optimization [7] Omar, M.R. and Abdulazeez, A., 2024. Fake news in social
algorithm. To make it more transparent and credible, network: A comprehensive review. The Indonesian Journal of
Computer Science, 13(3).
Explainable AI SHAP was used to explain the [8] Alghamdi, J., Luo, S. and Lin, Y., 2024. A comprehensive
classification of our proposed model. The survey on machine learning approaches for fake news
performance of our suggested model was measured detection. Multimedia Tools and Applications, 83(17),
based on accuracy, specificity, F1-score, precision, pp.51009-51067.
[9] Mridha, M.F., Keya, A.J., Hamid, M.A., Monowar, M.M. and
and recall, with superiority over other techniques like Rahman, M.S., 2021. A comprehensive review on fake news
BERT and XLNet. Our experimental outcomes detection with deep learning. IEEE access, 9, pp.156151-
revealed that the proposed RoBERTa with sequence 156170.
models performs better than the BERT 2.5% in [10] Roumeliotis, K.I., Tselikas, N.D. and Nasiopoulos, D.K.,
2025. Fake News Detection and Classification: A
accuracy and the XLNet model by 5.35% in accuracy. Comparative Study of Convolutional Neural Networks, Large
The model was tested with a single dataset that could Language Models, and Natural Language Processing
restrict the applicability of the findings across various Models. Future Internet, 17(1).
domains or types of review data. The future work [11] Hashmi, E., Yayilgan, S.Y., Yamin, M.M., Ali, S. and
Abomhara, M., 2024. Advancing fake news detection: Hybrid
should validate the model on more than one dataset deep learning with fast text and explainable ai. IEEE Access.
from diverse domains to examine its resilience. [12] Mohawesh, R., Salameh, H.B., Jararweh, Y., Alkhalaileh, M.
and Maqsood, S., 2024. Fake review detection using
REFERENCES transformer-based enhanced LSTM and
RoBERTa. International Journal of Cognitive Computing in
Engineering, 5, pp.250-258.
[1] Roumeliotis, K.I., Tselikas, N.D. and Nasiopoulos, D.K.,
[13] Dhiman, P., Kaur, A., Gupta, D., Juneja, S., Nauman, A. and
2025. Fake News Detection and Classification: A
Muhammad, G., 2024. GBERT: A hybrid deep learning
Comparative Study of Convolutional Neural Networks, Large
model based on GPT-BERT for fake news
Language Models, and Natural Language Processing
detection. Heliyon, 10(16).
Models. Future Internet, 17(1).
[14] Jamjoom, A.A., Karamti, H., Umer, M., Alsubai, S., Kim,
T.H. and Ashraf, I., 2024. Robertanet: Enhanced

Proceedings of WRFER International Conference, Pune, India, 28th June, 2025

54
 Ensemble Hybrid Deep Learning Framework for Fake News Detection Using RoBERTa and Explainable AI
robertatransformer based model for cyberbullying detection using sequential deep learning technique. Sensors, 22(18),
with glove features. IEEE access, 12, pp.58950-58959. p.6970.
[15] N. Abdelhakim Othman, D. S. Elzanfaly and M. M. M. [18] Umer, M., Alabdulqader, E.A., Alarfaj, A.A., Cascone, L.
Elhawary, "Arabic Fake News Detection Using Deep and Nappi, M., 2024. Cyberbullying Detection Using PCA
Learning," in IEEE Access, vol. 12, pp. 122363-122376, Extracted GLOVE Features and RoBERTaNet Transformer
2024, doi: 10.1109/ACCESS.2024.3451128. Learning Model. IEEE Transactions on Computational Social
[16] Sastrawan, I.K., Bayupati, I.P.A. and Arsa, D.M.S., 2022. Systems.
Detection of fake news using deep learning CNN–RNN based [19] Yang, X.S. and He, X., 2013. Firefly algorithm: recent
methods. ICT express, 8(3), pp.396-408. advances and applications. International journal of swarm
[17] Ali, A.M., Ghaleb, F.A., Al-Rimy, B.A.S., Alsolami, F.J. and intelligence, 1(1), pp.36-50.
Khan, A.I., 2022. Deep ensemble fake news detection model



Proceedings of WRFER International Conference, Pune, India, 28th June, 2025

55