𝗔𝗿𝘁𝗶𝗳𝗶𝗰𝗶𝗮𝗹 𝗜𝗻𝘁𝗲𝗹𝗹𝗶𝗴𝗲𝗻𝗰𝗲 𝗶𝗻 𝗣𝗵𝗮𝗿𝗺𝗮𝗰𝗲𝘂𝘁𝗶𝗰𝗮𝗹 𝗠𝗮𝗻𝘂𝗳𝗮𝗰𝘁𝘂𝗿𝗶𝗻𝗴: 𝗖𝘂𝗿𝗿𝗲𝗻𝘁 𝗜𝗻𝘀𝗶𝗴𝗵𝘁𝘀 𝗮𝗻𝗱 𝗙𝘂𝘁𝘂𝗿𝗲 𝗣𝗿𝗼𝘀𝗽𝗲𝗰𝘁𝘀

Pharma MES
2023 - Berlin
Artificial Intelligence in
Pharmaceutical Manufacturing:
Current Insights and Future
Prospects
Hernan Vilas
Thomas Halfmann
v1.1

Livinginthe
AIHype
• 350 billion words
• 175+ billion parameters
• 100+ million active users
• China now has at least
130 LLMs, accounting for
40% of the global total
and just behind the
United States' 50% share
(Source: Reuters)
• 21. Sep: Microsoft Copilot
announced
https://blogs.microsoft.com/blog/2023/09/21/announc
ing-microsoft-copilot-your-everyday-ai-companion/
2
Pharma MES 2023 – AI in Pharmaceutical Manufacturing: Current Insights and Future Prospects (Vilas & Halfmann)

GartnerDataand
AnalyticsHype
Generative AI is at the peak
of inflated expectations
3
Hype Cycle for Artificial Intelligence, 2023
Published 19 July 2023 (Gartner, Inc)
https://www.gartner.com/en/documents/4543699

The Reality of AI
4
Training data with
input/features x and
label/outputs y
corresponding to
each input. i.e:
images of cats and
dogs
Only input/features x, but
not corresponding
labels/outputs y. i.e: cluster
data that’s similar.
Seek to discover structure
in data
Might have a subset of data with
only inputs x and known
label/outputs y.
the machine learning task might
involve, searching the space, to find
some optimum. Reinforcement
learning: Alpha-Go example
There is a very large variety of machine learning methods and
algorithms, and a major challenge for implementing AI products, is
to be aware of this range of options and methods, and to match the
problem at hand, with the appropriate method.
Parametric
Non-Parametric
Parametric models are what we're most used
to, for most engineering and scientific
experiences. For example, in fitting a line or
curve, to some x data.
How models
learn from data
Model complexity depends on training
samples. i.e: decision trees, random forest,
SVM
Discriminative Given an input x, predict the probability of an
output y. Neural Networks. Deep,
Convolutional, Recurrent and Transformers
Generative Seek to model the joint probability of x and y, or
in some cases, the reverse probability. Given an
output y, what is the probability that different x
values are associated with that y?. GANs:
forward/reverse predictions. (deep-fakes)
Supervised
Learning
Unsupervised
Learning
Semi-
supervised
Learning

Finding a Way Forward
• Improve Human Performance:
AI can generate “super-human”
by augmenting our capabilities
• Science based AI will ensure
adoption in a regulated
environment like pharma.
• Democratize AI: Bring the user
friendliness of a chat bot into
the game to engage operators
and end users. LLMs can be
made available even for a small
company.
5

AIinDrugManufacturing–
TheRegulator’sView
FDA Discussion Paper
• FDA has recognized and embraced the potential of advanced
manufacturing and works collaboratively with companies to
support the use of advanced manufacturing
• FDA understands that AI may play a significant role in monitoring
and controlling advanced manufacturing processes
• FDA is asking the industry for feedback to be used for future policy
development
• Questions asked by FDA:
1. What types of AI applications do you envision …?
2. … aspects of the current regulatory framework …that should be considered by
FDA?
3. Would guidance in the area of AI in drug manufacturing be beneficial?
4. What are the necessary elements … to implement AI-based models in CGMP…?
5. What are common practices for validating and maintaining self-learning AI
models …?
6. What are the necessary mechanisms for managing the data used to generate AI
models in pharmaceutical manufacturing?
7. Are there other aspects …?
8. Are there aspects … not covered in this document that FDA should consider?
6
https://www.fda.gov/media/165743/download?attachment

AIapplicationsin
pharmaceutical
manufacturing
7
Process monitoring, fault
detection, anomalies and
trend monitoring
Predictive Maintenance,
calibration, validation
Advanced process controls -
Prediction of product
performance and quality.
PAT & Raman on AI models
Batch Release, process
deviation detection and
correction.
AI supported classification
of deviations in QMS
Self documenting CQV
reports for new equipment,
to significantly reduce the
time it takes to deploy new
machinery using generative
AI
Documentation preparation:
Master Batch Records
(MBRs) or population and
data analysis of the Product
Quality Annual Review
AI Co-pilot
SME / operator / expert
support: ‘error free
workflow’
Task automation
Digital Tech Transfer.
Process Design and Scale up
with AI-powered Digital
Twins. Virtual
Commissioning
AI Machine Learning Generative AI
https://www.biophorum.com/download/industry-feedback-on-artificial-intelligence-in-drug-manufacturing-fda-discussion-paper/

Challenges with AI-powered Solutions in Manufacturing
• Trust - Operators often distrust systems reliant on historical data during novel events. Building
confidence in AI requires change management.
à Change Management needed to ease the adoption of chatbot / auto-pilot mode for shop floor
operations.
• Bias remediation - Lifecycle management, responsible AI practices, audits and continuous
monitoring is needed to avoid any class specific bias (during training, deployment, algorithmic bias,
etc.)
• Explainability - Ability to explain in understandable terms how and why an AI model makes a certain
prediction or decision.
• Regulators require transparency into model logic and predictions.
• Complex AI models like deep neural networks are inherently black boxes.
• Connected Plant - Data Availability - Collecting large,
high-quality, labelled datasets required to train robust AI models
can be difficult due to proprietary concerns and siloed data.
à AI Hierarchy of Needs
8
https://hackernoon.com/the-ai-hierarchy-of-needs-18f111fcc007

Challenges with AI-powered Solutions in Manufacturing
• Validation - Extensive validation is required to prove GMP compliance.
Traditional statistical validation methods may not apply well to
AI systems.
• Regulatory Uncertainty - Lack of clear regulatory guidelines for
evaluating and certifying AI software creates ambiguity and risk for
manufacturers considering adoption.
à Leverage “Artificial Intelligence/Machine Learning (AI/ML)-Based
Software as a Medical Device (SaMD) Action Plan”
• Data privacy (IP protection, cybersecurity). Mainly with the extended
use of Generative AI platforms.
9
https://www.fda.gov/media/145022/download

ARealityCheck
Survey 2023:
AI in Biopharmaceutical
Manufacturing
-
How do MES solution
providers adopt AI today?
10

Survey: AI in Manufacturing
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
Is your company investing in developing AI as part of
your products?
Have you released any AI products or features?
Are the AI features fully integrated into the core
software?
11

Which areas of biopharmaceutical manufacturing will see the
most adoption in the next 3 years?
0% 10% 20% 30% 40% 50% 60% 70%
Advanced process control
Process monitoring, fault detection, anomalies and
trend monitoring
Expert support: ‘error free workflow’
Predictive maintenance
Production scheduling
Quality/defect detection
Yield optimization
Other (please specify)
12

What are the biggest potential benefits of AI in
biopharmaceutical manufacturing?
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
Automating complex tasks
Uncovering hidden insights
Increasing production efficiency
Reducing costs
Other (please specify)
13

How receptive are your biopharmaceutical clients to adopting
AI-based manufacturing software solutions?
0% 10% 20% 30% 40% 50% 60%
Very receptive
Somewhat receptive
Not very receptive
Not sure
14

How important are the following factors in your clients'
decisions to adopt AI? (5 being extremely important)
0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5
Demonstrated return on investment
Regulatory compliance
Data security
Algorithm transparency (explainability)
Ease of integration with existing systems
15

Survey: AI in Manufacturing (Observations)
16
1) MES solution providers look beyond the traditional AI solutions and benefits, e.g.
• Adaptive process control
• ChatGPT like product to support implementation partners, power users, operators, …
• Generative AI guided MBR design (Co-Pilot)
• Conversion of paper(-on-glass) MBRs into MES ready-to-use recipes
• Automatic text translation
• Plain text queries on manufacturing data
• …
2) Access to (the right) data is essential – the connected factory is a pre-requisite
3) Maturity and readiness is key: process, people and technology must be ready
and mature to adopt Generative AI
• Regulatory requirements must be fully understood and addressed
• An assessment based upon the Digital plant maturity model (DPMM) helps to understand and
create the roadmap for digital transformation and AI

17
2024Survey
AI in Biopharmaceutical
Manufacturing
• Maturity?
• Status of Adoption?
• Benefits of AI?
• Challenges, concerns &
risks?
• … Target audience:
• Industry
• Solution Providers
• Consulting & Services Companies

18
Conclusions …
• AI-powered solutions have reached the right level of maturity to support
decisions. In many cases, they are the only option to analyze and
evaluate highly complex data sets that are impossible for a human
• Data Strategy: capturing, integration and management of data assets will
ensure a consistent outcome of the AI models (garbage in / out principle)
• MES, Automation, OT Architecture are the foundations of the Factory of
the Future, powered by data, AI models and advanced process controls
to reach the vision of a continuous, fully automated and integrated
lights-off manufacturing
• Start now: understand where you are today, what is your digital maturity
(DPMM assessment) and where you want to go (the Roadmap to the
Factory of the Future)

We love to engage with you and stay connected …
Hernan Vilas Thomas Halfmann
19

20
Hernan Thomas

More Related Content

What's hot (20)

Recently uploaded (20)