Artificial General Intelligence (AGI) for the oil and gas industry:

a review
Jimmy Xuekai Li, Tiancheng Zhang, Yiran Zhu, and Zhongwei Chen*
School of Mechanical and Mining Engineering, The University of Queensland, Brisbane, QLD 4072, Australia
*Corresponding Author: [email protected]

Abstract
Artificial General Intelligence (AGI) is set to profoundly impact the oil and gas industry by introducing unprecedented
efficiencies and innovations. This paper explores AGI's foundational principles and its transformative applications,
particularly focusing on the advancements brought about by large language models (LLMs) and extensive computer
vision systems in the upstream sectors of the industry. The integration of Artificial Intelligence (AI) has already begun
reshaping the oil and gas landscape, offering enhancements in production optimization, downtime reduction, safety
improvements, and advancements in exploration and drilling techniques. These technologies streamline logistics,
minimize maintenance costs, automate monotonous tasks, refine decision-making processes, foster team collaboration,
and amplify profitability through error reduction and actionable insights extraction. Despite these advancements, the
deployment of AI technologies faces challenges, including the necessity for skilled professionals for implementation
and the limitations of model training on constrained datasets, which affects the models' adaptability across different
contexts. The advent of generative AI, exemplified by innovations like ChatGPT and the Segment Anything Model
(SAM), heralds a new era of high-density innovation. These developments highlight a shift towards natural language
interfaces and domain-knowledge-driven AI, promising more accessible and tailored solutions for the oil and gas
industry. This review articulates the vast potential AGI holds for tackling complex operational challenges within the
upstream oil and gas industry, requiring near-human levels of intelligence. We discussed the promising applications,
the hurdles of large-scale AGI model deployment, and the necessity for domain-specific knowledge in maximizing the
benefits of these technologies.

Keywords: Artificial General Intelligence (AGI), ChatGPT, Large Language Models (LLMs), Generative AI,
Multimodal, Oil and Gas Industry

1
1. Introduction
The oil and gas industry, a cornerstone of the global economy, is navigating a landscape marked by transformative
opportunities and unprecedented technical challenges (Prestidge, 2022). As the industry undergoes a pivotal shift
towards sustainable energy sources, it faces mounting pressure to enhance efficiency, reduce environmental impact,
and fulfill the growing energy demands of a global population (Hassan et al., 2024; Zohuri, 2023). This transition
requires navigating complex, high-risk operations across some of the planet's most challenging environments,
encompassing the exploration, development, production, storage, and abandonment (Epelle and Gerogiorgis, 2020;
Fagley, 2014; Islam and Khan, 2013; Speight, 2014).

1. Exploration
Geology and
Geophysics study for
hydrocarbon deposits

5. Abandonment 2. Development
Remove surface Drilling engineering,
facilities, Plug and seal Well construction,
the wells Formation evaluation

4. UG Storage 3. Production
Underground storage, Oil & gas extraction,
CCS, Hydrogen, Reservoir engineering,
Depleted reservoir Stimulation

Figure 1. Lifecycle and the critical stages of the upstream oil and gas industry. Artificial General Intelligence
(AGI) and its subsets generative AI, multimodal and especially the large language model (LLM) across each
stage would enhance efficiency and innovation in hydrocarbon exploration & production and underground
storage.

A general lifecycle of the oil and gas industry (Hussein, 2023; Tatjana, 2017) can be described by Figure 1. It delineates
the key stages in the upstream sector of the industry. The initial exploration phase demands not only a deep
understanding of geological structures but also the ability to accurately interpret subsurface data (Tearpock and Bischke,
2002; Wellmann and Caumon, 2018). Following exploration, the appraisal phase seeks to delineate the hydrocarbon
reservoir's potential (Amado, 2013). Development operations then take center stage, involving drilling engineering,
well construction, and comprehensive formation evaluation (Darling, 2005). Next, the production phase deploys
reservoir engineering techniques and advanced stimulation methods to optimize and maximize the recovery (Alvarado
and Manrique, 2010). This phase is pivotal in effectively managing the field's productive lifespan and extending the
viability of wells. At post-production, the focus might shift to innovative storage solutions depending on the reservoir
condition, reflecting the industry's commitment to resource management and environmental responsibility. Options
range from underground formations for hydrocarbon storage, such as gas storage facilities (Tarkowski, 2019), to carbon
capture and storage (CCS) (Bui et al., 2018) and hydrogen storage in depleted reservoirs (Muhammed et al., 2023). In
the final stages of a field's life, abandonment protocols are enacted where safety and environmental restoration take
precedence. This involves the careful removal of surface infrastructure and the sealing of wells, underscoring the
industry's dedication to sustainable practices (Babaleye et al., 2019).
In this complex landscape, the adoption of Artificial Intelligence (AI) has become more prevalent in each stage of the
oil and gas industry. These technologies have been used at the forefront of tackling the industry's challenges, offering
innovative solutions across the value chain. Some comprehensive review and summaries of the task specific or narrow
AI applications in the oil and gas industry can be found in these literatures (Hajizadeh, 2019; Pandey et al., 2020;
Sircar et al., 2021; Tariq et al., 2021; Waqar et al., 2023; Zhong et al., 2022).
The role of AI in this context has been proving crucial and cannot be overstated. This leads people to have more hope
for the on-going development of a new form of generalized AI or the Artificial General Intelligence (AGI) for the oil
and gas industry. From enhancing exploration success rates and drilling efficiency to advancing safety and sustainability,
AI and AGI are reshaping the future of oil and gas operations (Jacobs, 2024; Ma et al., 2023; Marlot et al., 2023;
Ogundare et al., 2023; Paroha and Chotrani, 2024; Waheed, 2023; Wang et al., 2023; Weijermars et al., 2023). As this

2
sector evolves in response to the shifting energy landscape, it leverages AI to not only address technical complexities
but also to pioneer a path toward a more sustainable and efficient future. This integration of cutting-edge technology is
a testament to the industry's resilience and forward-thinking ethos, promising to drive progress and innovation in the
years to come.
The rest of this paper is organised as follows. Section 2 discuss the foundations of AGI, including its history and
underlying technologies, such as LLMs, computer vision, and multimodal AI. Subsequent sections thoroughly examine
the deployment of these technologies within the industry, focusing on applications in data analysis, geophysics
interpretation, operational optimization, and beyond. The review highlights the latest innovative AGI applications that
are enhancing exploration, safety, and collaboration while also addressing the accompanying challenges and workforce
implications. It further outlook the future perspective of AGI in the oil and gas industry. The conclusion synthesizes
the review's main findings, emphasizing AGI's strategic importance and potential for addressing complex operational
challenges in the oil and gas industry.
Table 1. Comparison of Narrow AI and AGI.

Aspect Narrow AI AGI

Designed to perform specific tasks within a General cognitive abilities, can understand,
Definition and
limited range or context, such as learn, and apply intelligence to solve any
Scope
classification or image recognition systems. problem across a wide range of domains.
Learns from large amounts of labeled data Designed to learn and reason in various
Learning and
within its narrow scope, limited adaptability domains, adaptable to new scenarios,
Adaptability
to its defined domain. potentially with less data.
Less complex, widely used in various Far more complex, theoretical, and
Complexity and
industries, involves machine learning and experimental, aims to replicate human brain
Development
rule-based systems. processing.
Consciousness No self-awareness or consciousness, operates Theories often suggest potential for self-
and Self- purely within algorithmic constraints. awareness, although this is a subject of
Awareness debate.
Applications are vast but specific, such as Applications could be revolutionary, ranging
Applications spam filtering, recommendation systems, and from running corporations to conducting
autonomous vehicles. scientific research.
Data bias, need for large data sets, inability to Ethical considerations, the control problem,
Challenges generalize beyond trained tasks. technical difficulty in creating AGI systems.
Substantial impact in optimizing and Could revolutionize every aspect of human
Impact automating tasks, driving innovation within life but poses risks and uncertainties.
its scope.

2. Artificial General Intelligence (AGI) Background

AGI is an area of research and innovation aimed at developing computational systems capable of understanding,
learning, and performing any intellectual task that a human being can (Goertzel, 2014; Obaid, 2023; Youvan, 2024).
Unlike narrow or weak AI, which is designed to excel at a specific task or set of tasks, AGI seeks to achieve a versatile
and adaptive form of intelligence that mirrors human cognitive abilities. This includes the capacity for reasoning,
problem-solving, perception, and abstract thinking across diverse domains without needing task-specific programming.
Table 1 provides a simple comparison of the conventional narrow AI and the AGI based on summaries of multiple
literatures (Güner, 2022; Hoque, 2023; Kanade, 2022; Koc, 2023; VK, 2022).
The advancement of AI in the last decade, especially the recent couple of years (Figure 2), is an unprecedent
phenomenon primarily driven by the convergence of three key factors: the burgeoning availability of diverse and
voluminous training datasets (Pal et al., 2023), the evolution and enhancement of AI infrastructures (Gill et al., 2019;
Jeon et al., 2021), and the innovation within AI modelling techniques, notably Generative Adversarial Networks (GANs)
(Creswell et al., 2018; Goodfellow et al., 2014, 2020), diffusion models (Borji, 2022; Croitoru et al., 2023; Kingma et
al., 2021), and Transformers (Khan et al., 2022; Lin et al., 2022; Vaswani et al., 2017). These elements have all together
propelled AI capabilities forward, setting the stage for the rise of AGI. The discourse around AGI has permeated various
sectors, including agriculture (Lu et al., 2023), education (Latif et al., 2023), healthcare (Li et al., 2023b; Liu et al.,
2023a), and Internet of Things (IoT) (Dou et al., 2023), with its popularity surging post the introduction of OpenAI's
ChatGPT (Wu et al., 2023b). It's crucial to recognize that the ascendancy of AGI is not achieved by a single entity but
an outcome of collective work across academia and industry. A landmark in AGI's journey was the unveiling of the
Transformer architecture by Google in 2017 (Vaswani et al., 2017), a model designed to replace Recurrent Neural

3
Networks (RNNs) with self-attention mechanisms enabling sequential processing.
AGI's narrative took a dramatic turn with the advent of BERT (Bidirectional Encoder Representations from
Transformers) (Devlin et al., 2018), which, through its revolutionary pre-training on extensive corpora followed by
fine-tuning on specific tasks, set new benchmarks in natural language processing (NLP). The advent of BERT heralded
the dawn of foundation models (Koroteev, 2021), which shift the research focus from task-specific training to the
development of large-scale, pre-trained models that adapt to various applications through subsequent fine-tuning, few-
shot, or zero-shot learning strategies. This new wave has not only cemented the role of large-scale, task-agnostic models,
conventionally known as foundation models, within AGI but has also underscored their significance in domains beyond
NLP, like reinforcement learning (RL) (Wang et al., 2020) and computer vision (CV) (Khan et al., 2021).

Gemma Mixtral of Experts Q1

LLaMa2 Sora
GPT-4 StableLM3B 2024
GPT-3.5 Turbo
Grok-1 PanGu 3.0
Bard Gemini Q4 ImageFX
PaLM2
Hunyuan Imagen 2
MM1
LLaMa SenseNova PixArt-Σ
Q3 Sketch-to-Video
Claude3 Ernie Bot Tongyi Qianwen
Stable Video 3D
Leonardo
Q2
GPT Pilot Claude

Copilot SAM

Q1
ChatGPT Firefly
DALL-E3

OPT-IML Midjourney-V5 Runway

(Bing) Image Creator
2023 Ideogram
WebGPT Flan-PaLM
Stable studio Pika
FLAN GLM Q4
Ernie-ViLG Generative AI
InstructGPT
Github Copilot OPT

Alpha Code Q3
PanGu PaLM
DALL-E2
PanGu- Σ Dream studio DALL-E
Ernie Q2
Craiyon
Dream
Q1
2021 2022
Midjourney-V1
NightCafe

Figure 2. Recent developments of large language models (LLM, left side of the timeline curve) and multimodal
AI models (right side of the timeline curve), which combine the capabilities of LLM and CV (image, video, or
3D objects). Note: this figure does not intend to provide an exhaustive review or include all existing LLMs and
multimodal models. Instead, it aims to highlight some of the most renowned or commonly used LLMs and
multimodal models, which may also encompass APIs or web applications. The creation of this figure draws
inspiration from the Figure 3 in Zhao et al., (2023a), serving as a reference point rather than a comprehensive
catalogue.
For instance, visual foundation models, typified by DALL·E (Offert and Phan, 2022) and Imagen (DeepMind, 2023;
Gautam et al., 2024) and Sora (OpenAI, 2024), exemplify the migration of AGI into the visual domain, showcasing
adaptability across a multitude of tasks ranging from style transfer, image editing to image and video generation.
The AGI ecosystem is rapidly evolving to embrace multimodal foundation models capable of handling diverse data
forms, including text, images, video, and audio. Such models, exemplified by ChatGPT and GPT-4, facilitate cross-
modal knowledge transfer and unlock capabilities for multifaceted tasks requiring holistic data interpretation (Brin et
al., 2023; Egli, 2023).
Recent advances in large language models, like open source Grok-1 developed by X AI (XAI, 2024) trained with 314
billion parameters, have shown glimpses of AGI's potential, demonstrating a broad of understanding and reasoning that
approximates a more generalized form of intelligence. As such, these models are increasingly regarded as next
generation of mega scale of AGI, paving the way for an era where AI's applications are as equivalent as the human
intellect.

3. Towards AGI paradigm for the oil and gas industry

The journey towards embedding AGI in the oil and gas industry signifies a shift from conventional operational
frameworks to more intelligent, efficient, and sustainable practices. This section reviews the recent key applications of
LLM, LLM-based agents and multimodal, outlines the pathway toward realizing an AGI paradigm in the oil and gas
4
industry, highlighting strategic initiatives, technological advancements, and collaboration opportunities that pave the
way for this ambitious integration.
Adopting AGI in the oil and gas industry involves leveraging advanced AI capabilities to achieve comprehensive
cognitive functions enabling systems to perform a wide range of tasks autonomously (Weijermars et al., 2023). From
exploration and production to refining and distribution, AGI can offer unparalleled insights, predictive analytics, and
decision-making processes, leading to optimized operations and reduced environmental impact. However, transitioning
to an AGI-driven framework requires addressing considerable hurdles, including technological readiness, data
governance, workforce adaptation, and regulatory compliance.

3.1. Finetune LLM on domain knowledge

Although LLMs, such as GPT-4, LLaMA2, and Grok-1 models, have been trained through sophisticated methodologies
that leverage vast amounts of data and computational resources, finetuning LLMs on domain knowledge and private
data is a strategic approach to enhancing their utility, specificity, and effectiveness across various applications. It not
only ensures that the outputs are highly relevant and tailored to specific needs but also addresses concerns related to
data privacy, bias, and real-world applicability.
In the realm of geoscience, Deng et al., (2023) pioneers by introducing K2, the inaugural LLM specifically tailored for
geoscience, accompanied by an innovative suite of resources aimed at enhancing LLM research within this field. The
creation of K2 marks a major stride toward bridging the gap between advanced LLM capabilities and the intricate
requirements of geoscience research and application. This endeavour is supported by the development of GeoSignal, a
novel instruction-tuning dataset, including Natural Instructions (Mishra et al., 2021), the AI2 Reasoning Challenge
(Clark et al., 2018), Stanford Alpaca (Taori et al., 2023), and Dolly-15k (Conover et al., 2023), designed to refine LLM
responses to geoscience queries, and GeoBench, a benchmark for assessing LLM performance in geoscience contexts.
Deng et al., (2023) proposes a comprehensive adaptation strategy for pre-trained general-domain LLMs, i.e., LLaMA-
7B (Meta, 2023), to the geoscience domain, utilizing over 5.5 billion tokens from geoscience texts for further training
and fine-tuning the model with GeoSignal’s expert-curated data (Figure 3). This approach not only amplifies the
model's proficiency in geoscience knowledge but also equips K2 with tool-using capabilities, thus paving the way for
a multitude of geoscience applications. Through rigorous evaluations on GeoBench, the paper demonstrates the model's
enhanced understanding and utility in geoscience, setting a precedent for future research and applications in domain-
specific LLMs. The open-source release of K2 and its datasets underscores the collaborative spirit of this research,
inviting broader participation in the exploration and expansion of AGI in specialized geoscience domains.

Natural Wikipedia on
Instruction Geoscience

Dolly-15k Geoscience Knowledge

self-instruct
re-Structured
Task-specific Plain Text
Plain Text

GeoSignal Geoscience
Literature
.. K2 GeoLLaMA-7B LLaMA-7B
Instruction Tuning with Further Pretrain with
LoRA Geoscience curpus

Figure 3. The training pipeline for K2 comprises two key stages: the first involves further pre-training to
integrate geoscience knowledge, enhancing the model's expertise in this domain. The second stage, instruction
tuning with low-rank adaptation (LoRA) (Hu et al., 2021), is designed to fine-tune the model for human-like
interaction. This involves adapting the model to follow human instructions, emulate human responses, and align
more closely with human conversational patterns, ensuring a more natural and intuitive user experience (After
Deng et al., (2023)).

As a language model, K2 is capable of understanding geoscience materials and tailoring responses with suitable
prompts. As a generative model, K2 excels in creating detailed paragraphs and responding to queries with precise
answers, functioning similarly to a dynamic knowledge base that offers geoscientists professional assistance.
Furthermore, through the use of tools like Augment K2, the user can incorporate external information and
functionalities to produce reliable and effective outcomes. For instance, when posed with a question, K2 can initiate a
search using the GAKG's API, update its actions based on new data, and craft a prompt that leads to generating the

5
most appropriate answers. Figure 4 illustrates this process in action.

Figure 4. An example of tool augmented K2 for geoscience Q&A (After Deng et al., (2023)).
While K2 is optimized for geoscience, making it highly effective within this domain, K2's and its base model LLaMA-
7B's small scale (7 billion parameters in LLaMA-7B) limit its versatility compared to the more advanced LLM models
such as GPT-4, which boasts 1.76 trillion parameters and a Mixture of Experts (MoE) architecture (Betts, 2023; Shazeer
et al., 2017).
The MoE architecture is a scalable machine learning approach that dynamically allocates computational resources
across different "expert" models to efficiently process diverse tasks and data types, enhancing model performance and
adaptability. Thus, we suggest in the future work it should consider scaling up the model size for increased complexity
handling, incorporating a MoE architecture for improved efficiency and adaptability, and expanding domain-specific
training datasets with the latest geoscience data.

3.2. Enhancing Equipment Maintenance with LLM

The equipment maintenance strategies within the oil and gas sector have been enhanced through the application of
LLMs such as GPT-3 (Brown et al., 2020) and BERT (Devlin et al., 2018). LLMs analyse vast datasets, encompassing
operational parameters, historical performance data, and environmental factors. By understanding patterns and
predicting potential failures, these models enable a shift from traditional reactive maintenance to a more efficient
preventive approach. This not only helps in scheduling maintenance activities before the actual occurrence of failures
but also dramatically reduces unplanned downtimes and associated costs. Recognizing the sector's unique challenges,
including the management of vast unstructured data and industry-specific terminology, Abijith et al., (2023) tailors
these advanced NLP technologies to better understand and process maintenance-related texts. This initiative seeks to
unlock valuable insights from documents like manuals, work orders and technical records (Table 2), which are crucial
for optimizing both preventive and corrective maintenance operations and decision-making processes. Central to their
approach is the development of a custom LLM based on RoBERTa, a model chosen for its proven effectiveness in
various NLP tasks. The customization process involves extensive training on a corpus specifically gathered for this
purpose, including equipment manuals and plant maintenance records, to ensure the model's familiarity with the
industry's specific language. This process also includes the creation of a custom tokenizer, essential for accurately
segmenting text into analysable units, thus enabling the model to effectively decode the specialized vocabulary of the
oil and gas industry.
Table 2. Distribution of data used for training and tokenization in the custom LLM for equipment
maintenance (Abijith et al., 2023).
Data Set Number of words

Manuals 7.865 Million

Materials Data 1 Million

Notification Data 130 Million

Work Order Data 300 Million

Work Task Data 8.3 Million

It details the rigorous training regime, employing data parallel and distributed data parallel techniques to enhance

6
training efficiency and scalability (Zhao et al., 2023b). The model's performance is thoroughly evaluated through tasks
designed to reflect real-world applications, such as equipment segregation and masked word predictions. Despite facing
challenges, particularly in handling industry-specific abbreviations and terms, the findings illuminate the path for
considerable improvements, emphasizing the importance of comprehensive, domain-focused training datasets. By
suggesting ways to overcome identified limitations and proposing areas for further exploration, such as refining the
tokenization process and expanding the training datasets, the study encourages for a continued effort towards
integrating advanced NLP technologies into the sector. The proposed future work aims not only to refine the model's
performance but also to explore its application across a wider range of maintenance tasks, expected to markedly elevate
the efficiency and accuracy of maintenance strategies in the oil and gas industry.
Based on similar approach, Paroha and Chotrani, (2024) finetuned two advanced LLMs (i.e., TimeGPT and Time-
LLM), and applied them in predicting maintenance needs for Electrical Submersible Pumps (ESPs) for the oil
production. It was found that both models were effective in identifying key indicators of ESP health, which aligns with
established industry knowledge. TimeGPT (Figure 5) demonstrated slightly better performance overall, achieving
higher accuracy, precision, recall, and AUC-ROC values compared to Time-LLM. This demonstrates the potential of
these AI models to enhance predictive maintenance strategies within the sector. However, the integration of these AI
models into existing industrial setups poses substantial challenges, including the necessity for high-quality data and
compatibility with existing systems.

Figure 5. The architecture diagram of TimeGPT, (After Garza and Mergenthaler-Canseco, (2023)).

3.3. LLM as coding tool for oil and gas industry

LLMs, as part of generative AI, play a critical role in moving towards AGI by leveraging their ability to understand
and generate human-like text to assist with programming tasks, automation, and especially generating executable codes
based on natural language prompts so as to enhance various facets of software development, data handling, and
operational efficiency. These models are particularly adept at automating and improving tasks that traditionally require
considerable human intervention and expertise.
In the area of geosciences, Weijermars et al., (2023) recently integrate the LLM (ChatGPT) into the coding and data
processing of full-waveform inversion (FWI) for the seismic data, demonstrating generative AI's profound impact on
handling complex geoscientific data. They input a prompt to ChatGPT asking for a Python function based on Weiner
optimum filtering to perform spiking deconvolution on a series of seismic shot records. The AI efficiently returns a
function that handles the input parameters (e.g., seismic shot gather, length of the spiking filter, and pre-whitening
percentage) and produces the deconvolved seismic data (Figure 6). Rapid code generation by LLM reduces the time
that engineers and researchers spend on coding, allowing them to focus more on analysis and interpretation of results.
In addition, the generated code that is also syntactically correct and functionally effective, reducing the likelihood of
human error in manual coding.

7
 (a) Original seismic data (b) Processed seismic data from ChatGPT
generated programming

Figure 6. Comparison of Seismic Shot Gathers from a 2D Landline in East Texas, USA Panel: (a) shows the data
before applying spiking deconvolution, while panel (b) displays the results after deconvolution. The Python code
used for this process was generated by ChatGPT based on a prompt requesting a function for spiking
deconvolution using Weiner optimum filtering. This function processes seismic shot gathers, with inputs
including the gather itself, the length of the spiking filter, and the pre-whitening percentage. The output is the
deconvolved seismic data, demonstrating the effectiveness of the filtering approach (after Weijermars et al.,
(2023)).

Figure 7. Normalized power spectral density of the average trace before and after spiking deconvolution. The
spectral density is shown in blue prior to deconvolution and in red after the process. This comparison confirms
an enhancement in resolution following the application of spiking deconvolution code generated by ChatGPT
(after Weijermars et al., (2023)).
Figure 7 showcases the efficacy of the generated code through a graphical representation of the normalized power
spectral density (PSD) before and after spiking deconvolution. The comparison confirms resolution enhancement,
validating the LLM's capability in generating functional and reliable code that can be directly applied in geoscientific
research. While LLMs can generate code quickly, the responsibility remains on the researcher to verify the accuracy
and appropriateness of the output, especially in scenarios involving novel or complex data interpretations. Also
depending on specific research needs, the generated code may require adjustments or enhancements to better align with
unique project specifications or data characteristics.
Dhelie et al., (2023) also applied ChatGPT to oil and gas exploration tasks, highlighting successful implementations in
seismic image processing during pre-stack processing stages such as denoising, deblending, deghosting, and
debubbling. It emphasizes the ease of use and reductions in processing time achieved with AI tools such as ChatGPT
compared to traditional CPU-intensive methods. Figure 8 illustrates the use of the ChatGPT for generating
programming code for seismic image enhancements, noting its simplicity and efficiency in producing usable code with
minimal input.

8
 Figure 8. Seismic images enhancement by the programming code generated by the ChatGPT (After Dhelie et
al., (2023)).
In sum, the use of generative AI in generating code for data processing in geosciences and all other sectors in the oil
and gas industry represents a big leap forward in operation and research productivity and capability. As AGI
technologies continue to evolve, their integration into petroleum engineering and geoscientific research is expected to
deepen, offering even more sophisticated tools for data analysis and interpretation.

3.4. Multimodal AI on enhancing real-time production prediction

Multimodal AI refers to the models that integrate and process two or more different types of data modalities (such as
text and images), reflecting an important step towards AGI. These systems are designed to leverage the complementary
information available from different modalities to improve the accuracy and robustness of AI applications. Its ability
to integrate diverse data types mirrors the versatility and adaptability required for AGI, marking a great advancement
over narrow AI systems.
The upstream oil and gas industry are characterized by the vast and varied types of data it generates, including seismic
data, drilling logs, core samples, and production data among others. This diversity of data spans numerous modalities,
such as numerical data, text, and images, making multimodal AI particularly important for this sector. One noteworthy
development is the integration of multi-modal data into predictive models for real-time oil production. Jiang et al.,
(2023) use an improved Artificial Fish Swarm Algorithm (AFSA) combined with a Long Short-Term Memory (LSTM)
network to optimize production prediction models using multimodal data containing image features and production
data.
The core of the proposed methodology is the fusion of different data modalities to improve the predictive accuracy of
oil production models. The study uses image features extracted from indicator diagrams, which are traditional tools for
monitoring equipment and production processes (Figure 9. These image features are combined with time-series
production data, creating a rich, multimodal dataset that provides a comprehensive view of the production dynamics.
Measured indicator diagram Image after binarization Image after edge detection Image after edge refinement

Figure 9. Feature extraction on indicator diagrams. The indicator diagram often contains rich information, e.g.,
load, displacement scale, coordinate axis, and so on, but it cannot be used directly for production prediction.
The features extractions (after process e.g., binarization, edge detection and refinement) can be deemed as the
features of the indicator diagram and used for production prediction (Modified from Li et al., (2024)).
The LSTM can process time series data effectively, while the AFSA is utilized to optimize the LSTM's hyper-
parameters dynamically. This combination addresses two major challenges in traditional predictive modelling in the
oil and gas industry: the reliance on single-modal data which limits predictive performance, and the manual setting of
model parameters, which impedes achieving the optimal model configuration.
The effectiveness of the proposed model is demonstrated through experiments, which show that the AFSA-LSTM
model achieves a mean absolute percentage error (MAPE) of just 4.313%, outperforming both traditional methods and
other deep learning models. The integration of multimodal data not only enhanced the model's accuracy but also its
robustness against varying operational conditions.
This study demonstrates the potential of multimodal AI applications in revolutionizing production prediction in the oil
and gas industry. By integrating diverse data types and employing advanced machine learning techniques, operators
9
can achieve more accurate predictions, leading to better resource management and optimized operational decisions.

3.5. Multimodal AI on enhancing standard knowledge extraction

The oil and gas industry have been generating huge amount of multimodal data during the complex and intensive
operations, ranging from texts and documents to images and videos. Traditional methods, which often rely on manual
processing of single-modal data, are inadequate for the semantic representation and integration of the industry's
multimodal data. Multimodal AI could play a pivotal role to enhance the extraction and organization of knowledge
from the vast and diverse data in the industry, which are often unstructured or semi-structured.
A recent study by Huang et al., (2024) explores an advanced approach for automatic extraction of multimodal
knowledge specifically tailored for the standard documents in petroleum field. The algorithm designed to extract
fragmented standard indicator attributes primarily involves two components: table extraction and paragraph extraction.
Both modules rely on a part-of-speech classifier. Through table extraction, the user can derive ontology and indicator
items directly from tables. Similarly, paragraph extraction allows the user to capture ontology from the applicable scope
paragraphs in standard documents. The outputs from both extraction methods are then merged and incorporated into
the indicator database. Figure 10 presents a flowchart illustrating the process of the standard indicator extraction
algorithm.
The new multimodal AI integrates various AI technologies, including text clustering, keyword extraction, and advanced
classification techniques, to handle the multimodal nature of the data effectively. For instance, the study enhances text
classification systems to automatically categorize documents and assigns them to predefined categories, facilitating
easier management and retrieval of information. The approach not only utilizes traditional statistical methods but also
incorporates rule-based classification technologies, allowing for dynamic adjustment according to specific industry
needs.

Figure 10. Flowchart of the standard indicator extraction algorithm (After Huang et al., (2024)).
The application of such multimodal AI systems enables the automatic organization and semantic analysis of vast
quantities of data, which not only improves efficiency but also enhances the accuracy and accessibility of information.
This ability to quickly and accurately extract and organize standard knowledge would also support the industry's
ongoing efforts in innovation and regulatory compliance. By effectively leveraging the power of AI to process and
analyse multimodal data, the industry can expect substantial advancements in how information is managed and utilized.

3.6. Multimodal AI on gas leakage detection

An important application of multimodal AI systems in the industry is in the robust sensor fusion, which refers to the
process of combining sensory data from disparate sources to produce more accurate, consistent, and useful information
than that provided by any individual sensor alone so as to capture a comprehensive understanding of the environment.
A recent study by Rahate et al., (2023) demonstrates the application of multimodal co-learning to evaluate the
robustness of sensor fusion in gas detection systems, which is crucial for ensuring safety in various environments across

10
the industry. The research focuses on using a combination of gas sensor data and thermal images to enhance the
detection and identification processes (Figure 11).
It introduces a methodological innovation by employing multimodal co-learning, which can deal with scenarios where
data from one or more sources may be noisy or unavailable. The multimodal approach integrates data from primary
gas sensors and thermal cameras to form a robust detection system. This integration allows the system to compensate
for missing or noisy data from one modality by leveraging information from the other, ensuring reliable operation under
varying conditions. The experimental results confirm that multimodal sensor fusion can enhance the system's
robustness compared to traditional single-source methods. For instance, the addition of low-resolution thermal modality
data aids in compensating for up to 20% missing sensor data and 90% missing thermal image data. This capability is
crucial in real-world industrial settings, where sensor outputs can be unpredictable and affected by environmental
factors.

Figure 11. Multimodal co-learning system architecture for evaluating the robustness of sensor fusion (After
Rahate et al., (2023))
In a later study, Attallah, (2023) introduces novel multimodal data fusion strategies termed "intermediate" and
"multitask" fusion, which are more sophisticated compared to the multimodal co-learning focus in Rahate et al.,
(2023)'s work. Intermediate fusion involves integrating features at a mid-level using techniques like Discrete Wavelet
Transform (DWT), improving the handling of spatial, spectral, and temporal information. Multitask fusion, on the other
hand, employs Discrete Cosine Transform (DCT) for dimensionality reduction and feature integration from different
AI architectures, providing a robust way to handle diverse data types efficiently. The system achieves high detection
accuracies (98.47% and 99.25% for intermediate and multitask fusion respectively), which is a remarkable
improvement over previous methodologies.

3.7. Zero-shot learning on drill core images

Zero-shot learning (ZSL) is a machine learning technique where a model learns to recognize objects, concepts, or
classes that it has not seen during training. It relies on understanding general characteristics and properties that can be
transferred from seen classes to unseen classes. This ability is a step forward towards creating more flexible and
general-purpose models, which is one of the foundational goals in the pursuit of AGI.
Drill cores are invaluable in providing direct subsurface information and continuous record of the rock formation. The
accurate measurement of fracture frequency or fracture density from drill core samples thus provides vital information
that impacts the exploration and exploitation of natural resources, the assessment of geological hazards, and the design
and construction of underground structures. The AI models have exhibited superior performance and efficiency
compared to traditional manual core logging methods, offering an automated approach to geological and geotechnical
analysis, which represents a technological shift aimed at enhancing the precision and efficiency of geological
explorations. (Li et al., 2023a; Su et al., 2023). We tested different computer vision AI models on the core sample mask
detection and classification task. Figure 12 offers a comparative analysis of core sample instance segmentation with
overlaid segmentation masks, applying three distinct AI models: Mask R-CNN, Mask2Former, and Segment Anything
(SAM).
Mask R-CNN (He et al., 2017) relies on labelled data and utilizes transfer learning from pre-trained model to perform
instance segmentation. Here, the precision of Mask R-CNN is evident as it can perform segmentation on the images of
core samples with accuracy, but the requirement for extensive labelled datasets and the associated preparatory work
limits its scalability and adaptability to new, unlabelled datasets. The more recent model Mask2Former (Cheng et al.,
2022) refines the capabilities of its predecessor through the integration of transformer-based architectures, enabling a
more sophisticated understanding of the global context within the image. Its performance, potentially surpassing Mask
R-CNN, is indicative of advancements in AI, however, it still faces the challenges of data dependency and the need for
a substantial annotation effort.
The latest Segment Anything (SAM) model (Kirillov et al., 2023) displays what is referred to as a 'zero-shot prediction',
which does not require labelled data to predict and segment new instances. It operates on the principles of zero-shot

11
learning, where the model generalizes from seen to unseen data, showcasing an approach that aligns with the trends
towards AGI. This generalization capability is especially crucial in fields where labelling data is impractical due to the
variety and complexity of instances, such as core samples with diverse geological features.
The generalization of SAM not only underscores the potential for AI models to progress towards AGI but also
demonstrates a remarkable evolution in instance segmentation technology. The transition from reliance on labelled data
to the ability to predict without prior direct examples illustrates a growing autonomy in AI systems, enhancing their
applicability in real-world scenarios and reducing the time and resources required for deployment.

(a) Mask RCNN (2017) model result

(using labelled data and transfer learning)

(b) Mask2Former (2022) model result

(using labelled data and transfer learning)

(c) Segment Anything (2023) model result

(zero shot blind prediction)

Figure 12. Comparative analysis of core sample instance segmentation (masks over image) using three AI models:
Mask R-CNN, Mask2Former, and Segment Anything (SAM), highlighting their respective accuracies in core
sample identification.

3.8. Zero-shot learning on digital rock physics

The need for high-accuracy segmentation in reservoir modelling has also been met with the powerful, zero-shot
learning SAM model, demonstrating an ability to segment images interactively and automatically without extensive
labelled datasets. Traditional semantic segmentation models relied heavily on large-scale annotated datasets, making
them time-consuming to prepare, especially for complex CT and SEM rock images. SAM improves the segmentation
process by delivering a zero-shot segmentation capability, which is important for digital rock physics with its limited
data and intricate image features. Ma et al., (2023) has addressed the zero-shot digital rock image segmentation with a
fine-tuned SAM model. Initially, the model encodes images and prompts into high-dimensional vectors using a Vision
Transformer and a text encoder, respectively. These vectors are then merged and processed by a mask decoder to
produce precise segmentation masks corresponding to the prompts. The versatility of SAM is highlighted through its
support for various prompt types—points, boxes, or text—each facilitating different segmentation strategies. Text
prompts allow for the segmentation of objects based on abstract concepts, while points and boxes guide the model
spatially, enhancing the accuracy of the segmentation process. The SAM model, however, faced limitations when
applied to digital rock images due to their complexity and low-contrast features. To address this, the model underwent
fine-tuning, resulting in the RockSAM variant, which improved segmentation accuracy for digital rock images without
sacrificing its zero-shot learning capabilities (Figure 13). This adaptation ensures RockSAM's effectiveness even on
resource-constrained devices, presenting a valuable tool for digital rock image analysis.
The RockSAM has shown remarkable efficiency in generating high-quality segmentation masks, overcoming the need
for detailed training or complex labelling. Its generalized approach and minimal data requirements stand in contrast to
previous models, which, despite their advancements, still needed labelled data and extensive transfer learning. Thus,
RockSAM represents a more evolved AI that aligns with the principles of AGI—learning and adapting with minimal
human intervention and data. It not only elevates the accuracy of digital rock image analysis but also suggests a future
where AGI can seamlessly manage other complex tasks in the oil and gas sector. This trend towards AGI-equipped
models suggests an upcoming era of more autonomous, efficient, and versatile AI applications in the industry. The
success of RockSAM advances this movement, demonstrating that even with the complex and nuanced task of rock
image segmentation, AI can learn to excel with limited datasets. As AI continues to advance, we anticipate models
becoming even more adept at handling the challenges and complexities of the natural resource sector.

12
Figure 13. RockSAM, fine-tuned the SAM model for digital rock images, modified the weights in the mask
decoder, while maintaining the other components as they are. This process initiates with the transformation of
the image into a high-dimensional vector, capturing its intricate details. The prompt (points, a box, or text) is
converted into a unique vector and merged into the mask decoder (After (Ma et al., 2023)).

3.9. Summary
This section discusses the implementations of AGI within the oil and gas industry, highlighting the transition from
conventional operational frameworks to more intelligent, efficient, and sustainable practices. The section reviews key
applications of large language models (LLMs), LLM-based agents, and multimodal systems, underscoring the
pathways toward realizing an AGI application in the industry. It demonstrates how adopting AGI can bring about
comprehensive cognitive functions, enabling systems to perform a wide range of tasks autonomously. This includes
optimizing operations and reducing environmental impact.
Specific advancements and initiatives are outlined, including the strategic finetuning of LLMs on domain knowledge,
enhancing equipment maintenance through AI, and using multimodal AI to improve real-time production prediction.
Each of these points is discussed with examples and case studies that illustrate the potential benefits and challenges
associated with integrating AGI technologies into various facets of the oil and gas industry.

4. Challenges and future perspectives: shift towards agent-oriented models

Large Language Models (LLMs) and their multimodal form advanced with vision perception capability excel in
processing and responding to complex queries, thanks to their sophisticated reasoning and computer vision capabilities.
LLMs are adept at breaking down intricate problems into manageable steps, providing comprehensive solutions,
actions, and evaluations step-by-step. This capability enables them to address a broad spectrum of challenges effectively.
However, despite these capabilities, LLMs are inherently limited as self-contained systems. They cannot access up-to-
date data or draw upon specialized knowledge specific to particular fields, which can result in the production of
inaccurate information or "hallucinations." Traditional methods of fine-tuning pre-trained LLMs involve specific
adjustments to the neural network weights and datasets, which may restrict their general applicability. Furthermore,
LLMs typically face challenges with tasks that require precise numerical computations or staying updated with the
latest developments.
To mitigate these limitations, there is a growing recognition of the necessity to integrate external data sources and
additional tools. Thus, developing autonomous agents that work in conjunction with LLMs is becoming increasingly
crucial (Xi et al., 2023). These agents enhance the functionality of LLMs by enabling real-time data integration and
application-specific adaptability (Guo et al., 2024; Wang et al., 2024).
Due to the reasons stated above, since the second half of 2023, there has been a redirection in the strategic focus of
leading technology firms and academics, with a move towards the development and implementation of agent-based
systems. OpenAI API allows the user to call the core LLM functions to build their own smart agent, reflecting a broader
industry trend (GPT-5, 2023). Concurrently, Amazon announced the launch of Amazon Bedrock Agents, augmenting

13
 their AI capabilities with advanced agent functionalities designed to handle more sophisticated tasks autonomously
(AWS, 2023). Google developed the Vertex AI Agent Builder, which allows developers build and deploy gen AI
experiences using natural language or open-source frameworks like LangChain (Google, 2024). These developments
are indicative of an industry-wide toward agent-based applications, which was presaged by earlier projects such as
Stanford’s "AI Town," demonstrating the practical integration of AI agents in simulated environments (Park et al.,
2023). Liu et al., (2023b) released AgentBench, an AI agent capability evaluation tool, which provides a standardized
framework for assessing the performance and capabilities of AI agents. Ruan et al., (2023) proposes a structured
framework for LLM-based AI agents, introduces two types of agents for executing tasks, and evaluates their abilities
in Task Planning and Tool Usage (TPTU), aiming to enhance the application of LLMs in complex real-world scenarios
and identify areas for further research and development.
All these technological advancements and the shift towards AI agents would hold profound implications for the oil and
gas industry. These agents offer the potential to revolutionize industry operations by enabling real-time data integration,
automating complex decision-making processes, and potentially overseeing complete operational workflows
autonomously. The alignment of these advanced agents with the industry’s digital transformation goals suggests a future
where AI agents not only augment but centrally manage critical industry operations. As depicted in Figure 14, we
would introduce the LLM-based agent structure on a geosteering drilling example, where the AI-driven agent assists
in the complex decision-making process of automatic geo-steering, a technology used in directional drilling to navigate
through the subsurface to the target zone.

Environment Perception
Inputs

We're monitoring the real-time LWD

data. Based on current trends and
geological models, should we adjust
our drilling parameters to stay within
the pay zone?" Tokenization
Reasoning
Storage
Memory Knowledge
[Perception] Analyzing the latest
real-time logs in conjunction with
our geo-steering model [Reasoning]
indicates an impending deviation Generalize/
beyond the optimal drilling window. Agent LLM Transfer  Summary  Recall
[Reasoning] To mitigate this, I  Retrieve  Learn
recommend an immediate adjustment
Planning Inference
of the drilling parameters.
[Tools/Embodiment] Specifically,
altering the weight on bit and the
rotation speed should correct our
trajectory. I will also adjust the Action Decision-Making
downhole tool's position to maintain
our course within the reservoir Text Tools
boundaries." Calling API …
…

Embodiment

Figure 14. Enhancing Drilling Precision with LLM-Based Smart Agents in Geo-Steering Operations
In the context of real-time wellbore positioning, the smart agent's role extends from interpreting live Logging While
Drilling (LWD) data to making informed decisions about drill path adjustments. The agent functions as a cognitive
intermediary, analysing incoming data streams against geological models to ensure that the drill bit stays within the
optimal pay zone. This process is depicted in Figure 14 through various stages:
(i) Perception: The smart agent gathers diverse inputs from environmental sensors, LWD tools, and geological
databases, processing this data to maintain an up-to-date perception of the drilling environment.
(ii) Reasoning: Leveraging its extensive knowledge base, the agent applies reasoning to understand the implications
of the data. For instance, it can predict the trajectory of the drill bit and anticipate the need for adjustments to stay
within the reservoir boundaries. The agent also plans actions by considering the current trajectory and the desired
outcome. It simulates various scenarios to identify the most effective drilling parameters that align with the
geological objectives.

14
(iii) Action: The agent then executes the plan by adjusting the drilling parameters, such as the weight on bit and rotation
speed, to correct the drill path, thus embodying the decision-making process in the physical environment of the
drilling operation.
The LLM-based smart agent is equipped with advanced cognitive capabilities, simulating a level of understanding and
problem-solving that mimics human expertise. The model's ability to interact with text-based prompts enables it to
interpret instructions and provide recommendations that are comparable with the expertise of seasoned drillers. The
agent's ability to process large volumes of data and make real-time decisions reduces operational downtime and
optimizes resource extraction.
Real-world situations frequently require collaborative efforts to achieve optimal task execution. Reflecting the
principles of collective human behaviour, researchers have developed a multi-agent framework (Devi et al., 2024; Wu
et al., 2023a). This framework is designed to operate collectively, enabling a group of agents to supass the capabilities
of individual agents working alone. A multi-agent system (MAS) is a system made up of multiple LLM-based agents
that interact with each other in a shared environment (Wu et al., 2023a). These agents are intelligent in the sense that
they can make decisions and take actions autonomously, and they can also communicate with each other to achieve a
common goal or individual goals that may conflict (Hong et al., 2023; Oroojlooy and Hajinezhad, 2023; Orr and Dutta,
2023). Thus, the MAS are very useful for solving problems that are too complex for a single agent to handle on its own,
such as drilling engineering.
The Figure 15 showcases a network of multiple agents (Agent 1 to Agent 6, each agent has unique expertise besides
the general knowledge in the foundation LLM model, such as geology or drilling engineering), highlighting the
potential for collaborative decision-making or distributed tasks across various operational sectors. This ensemble
illustrates the interconnectedness of the agents, working in unison through complex communication pathways that
allow for real-time collaboration and decision-making. The inter-agent dynamics enable the system to adapt
dynamically, leveraging the collective intelligence of the group to solve problems more efficiently than any agent could
independently.

Agent Environment

Action

 Capability
 Personality

Agent 2 Agent 3
Virtual Env or Real Env

Agent 1 Agent 4

Agent 1agent
Individual

Human & Resources

Agent 6 Agent 5
Perception
Agent group

Figure 15. Multi-Agent Collaboration Framework Enhancing Geo-Steering Drilling Operations (Modified
from Xi et al., (2023))
The agents interact with the "Environment," which is illustrated in two domains: the virtual (e.g., subsurface geological
model and well plan) and the real (e.g., hardware, tools on a drilling rig). This dual representation indicates the system's
flexibility in operating within simulated models—useful for planning and forecasting—as well as its capacity to act in
the physical world, directly influencing drilling operations on an oil rig. Actions taken by the agents are reflected in
the environment, influencing factors such as drilling trajectory and equipment adjustments. Simultaneously, the agents
receive feedback from the environment ("Reception"), which they use to refine their future actions, thus creating a
feedback loop that enhances the operation's precision. The "Human" and "Resources" within the environment suggests
the integration of human expertise and available materials and data into the decision-making process. This human
intervention ensures that the automated system remains aligned with human oversight and industry standards,
15
optimizing the use of available resources.

5. Conclusions
The integration of Artificial General Intelligence (AGI) in the oil and gas industry marks the beginning of a
transformative era characterized by the deployment of intelligent systems capable of performing complex tasks with
near-human intelligence. This review paper has underscored the strategic importance of AGI's innovative applications
and its potential to tackle operational challenges within the upstream sectors of the industry. Through Large Language
Models (LLMs) and advanced computer vision systems, AGI has already started to enhance exploration rates, drilling
efficiency, safety, and sustainability.
Despite the promising future AGI signals, the industry faces a lot of challenges in adopting these technologies on a
large scale. Among these are the needs for skilled professionals for effective implementation, limitations of model
training on constrained datasets affecting adaptability, and the essential integration of domain-specific knowledge for
maximizing the benefits of AGI technologies. As the technology progress, the move towards agent-oriented models
(LLM-based agent) presents an opportunity to redefine the landscape of industry operations. The smart agents, built on
the foundations of AGI, are set to play an important role in automating decision-making processes, reducing operational
downtimes, and optimizing resource extraction. The successful application of the Segment Anything Model (SAM) in
zero-shot learning tasks, such as drill core image analysis and digital rock physics, demonstrates the strides AGI has
made in achieving operational efficiency.
The trajectory towards AGI is clear, with a focus on refining the collaborative and autonomous functions of multi-agent
systems. These systems promise enhanced precision in complex environments, like those encountered in geosteering
drilling operations, where real-time data interpretation and adaptive decision-making are crucial.
This review has articulated the vast potential AGI holds and the pathway toward its realization in the oil and gas industry.
As the technology continues to advance, it is expected to bring forth an era of more autonomous, efficient, and versatile
AI applications, driving innovation and productivity in the sector.

6. Declaration of the use of generative AI

The authors acknowledge the use of ChatGPT, a large language model developed by OpenAI, and Grammarly, a writing
and grammar-checking software, in preparing this article. They were used to improve the writing quality of this paper.

References

Abijith, P., P. Patidar, G. Nair, and R. Pandya. 2023, Large Language Models Trained on Equipment Maintenance Text.
Paper read at Abu Dhabi International Petroleum Exhibition and Conference.
Alvarado, V., and E. Manrique, 2010, Enhanced oil recovery: field planning and development strategies: Gulf
Professional Publishing.
Amado, L., 2013, Reservoir exploration and appraisal: Gulf Professional Publishing.
Attallah, O., 2023, Multitask Deep Learning-Based Pipeline for Gas Leakage Detection via E-Nose and Thermal
Imaging Multimodal Fusion: Chemosensors, 11, no. 7, 364.
AWS, A., 2023, Agents for Amazon Bedrock. https://aws.amazon.com/bedrock/agents/.
Babaleye, A. O., R. E. Kurt, and F. Khan, 2019, Safety analysis of plugging and abandonment of oil and gas wells in
uncertain conditions with limited data: Reliability Engineering & System Safety, 188, 133-141.
Betts, S., 2023, Peering Inside GPT-4: Understanding Its Mixture of Experts (MoE) Architecture.
https://medium.com/@seanbetts/peering-inside-gpt-4-understanding-its-mixture-of-experts-moe-architecture-
2a42eb8bdcb3.
Borji, A., 2022, Generated faces in the wild: Quantitative comparison of stable diffusion, midjourney and dall-e 2:
arXiv preprint arXiv:2210.00586.
Brin, D., V. Sorin, A. Vaid, A. Soroush, B. S. Glicksberg, A. W. Charney, G. Nadkarni, and E. Klang, 2023, Comparing
ChatGPT and GPT-4 performance in USMLE soft skill assessments: Scientific reports, 13, no. 1, 16492.
Brown, T., B. Mann, N. Ryder, M. Subbiah, J. D. Kaplan, P. Dhariwal, A. Neelakantan, P. Shyam, G. Sastry, and A.
Askell, 2020, Language models are few-shot learners: Advances in neural information processing systems, 33,
1877-1901.
Bui, M., C. S. Adjiman, A. Bardow, E. J. Anthony, A. Boston, S. Brown, P. S. Fennell, S. Fuss, A. Galindo, and L. A.
Hackett, 2018, Carbon capture and storage (CCS): the way forward: Energy & Environmental Science, 11, no.
16
 5, 1062-1176.
Cheng, B., I. Misra, A. G. Schwing, A. Kirillov, and R. Girdhar. 2022, Masked-attention mask transformer for universal
image segmentation. Paper read at Proceedings of the IEEE/CVF conference on computer vision and pattern
recognition.
Clark, P., I. Cowhey, O. Etzioni, T. Khot, A. Sabharwal, C. Schoenick, and O. Tafjord, 2018, Think you have solved
question answering? try arc, the ai2 reasoning challenge: arXiv preprint arXiv:1803.05457.
Conover, M., M. Hayes, A. Mathur, X. Meng, J. Xie, J. Wan, A. Ghodsi, P. Wendell, and M. Zaharia, 2023, Hello Dolly:
Democratizing the magic of ChatGPT with open models: Databricks blog. March, 24.
Creswell, A., T. White, V. Dumoulin, K. Arulkumaran, B. Sengupta, and A. A. Bharath, 2018, Generative adversarial
networks: An overview: IEEE signal processing magazine, 35, no. 1, 53-65.
Croitoru, F.-A., V. Hondru, R. T. Ionescu, and M. Shah, 2023, Diffusion models in vision: A survey: IEEE transactions
on pattern analysis and machine intelligence.
Darling, T., 2005, Well logging and formation evaluation: Elsevier.
DeepMind, G., 2023, Imagen 2. https://deepmind.google/technologies/imagen-2/.
Deng, C., T. Zhang, Z. He, Q. Chen, Y. Shi, L. Zhou, L. Fu, W. Zhang, X. Wang, and C. Zhou, 2023, K2: A Foundation
Language Model for Geoscience Knowledge Understanding and Utilization: arXiv preprint arXiv:2306.05064.
Devi, K. V. R., B. Smitha, S. Lakhanpal, R. Kalra, V. A. Sethi, and S. K. Thajil. 2024, A review: Swarm Robotics:
Cooperative Control in Multi-Agent Systems. Paper read at E3S Web of Conferences.
Devlin, J., M.-W. Chang, K. Lee, and K. Toutanova, 2018, Bert: Pre-training of deep bidirectional transformers for
language understanding: arXiv preprint arXiv:1810.04805.
Dhelie, P., A. Evensen, and A. Bugge. 2023, Increasing your exploration success using AI, ML and ChatGPT. Paper
read at 84th EAGE Annual Conference & Exhibition.
Dou, F., J. Ye, G. Yuan, Q. Lu, W. Niu, H. Sun, L. Guan, G. Lu, G. Mai, and N. Liu, 2023, Towards Artificial General
Intelligence (AGI) in the Internet of Things (IoT): Opportunities and Challenges: arXiv preprint
arXiv:2309.07438.
Egli, A., 2023, ChatGPT, GPT-4, and other large language models: The next revolution for clinical microbiology?:
Clinical Infectious Diseases, 77, no. 9, 1322-1328.
Epelle, E. I., and D. I. Gerogiorgis, 2020, A review of technological advances and open challenges for oil and gas
drilling systems engineering: AIChE Journal, 66, no. 4, e16842.
Fagley, E. M., 2014, Offshore Energy Geopolitics: An Examination of Emerging Risks to Future Oil and Gas Activities
in Hotspot Maritime Regions, Johns Hopkins University.
Garza, A., and M. Mergenthaler-Canseco, 2023, TimeGPT-1: arXiv preprint arXiv:2310.03589.
Gautam, S., P. N. Venkit, and S. Ghosh, 2024, From Melting Pots to Misrepresentations: Exploring Harms in Generative
AI: arXiv preprint arXiv:2403.10776.
Gill, S. S., S. Tuli, M. Xu, I. Singh, K. V. Singh, D. Lindsay, S. Tuli, D. Smirnova, M. Singh, and U. Jain, 2019,
Transformative effects of IoT, Blockchain and Artificial Intelligence on cloud computing: Evolution, vision,
trends and open challenges: Internet of Things, 8, 100118.
Goertzel, B., 2014, Artificial general intelligence: concept, state of the art, and future prospects: Journal of Artificial
General Intelligence, 5, no. 1, 1-48.
Goodfellow, I., J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y. Bengio, 2014,
Generative adversarial nets: Advances in neural information processing systems, 27.
Goodfellow, I., J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y. Bengio, 2020,
Generative adversarial networks: Communications of the ACM, 63, no. 11, 139-144.
Google, 2024, All 218 things we announced at Google Cloud Next ’24 – a recap.
https://cloud.google.com/blog/topics/google-cloud-next/google-cloud-next-2024-wrap-up.
GPT-5, 2023, Building an AI Agent with OpenAI’s Assistants API: A Comprehensive Guide.
https://medium.com/@trendingAI/building-an-ai-agent-with-openais-assistants-api-a-comprehensive-guide-
17
 c23846221750.
Güner, A., 2022, General AI vs. Narrow AI : 2022 Guide. https://www.plugger.ai/blog/general-ai-vs-narrow-ai-2022-
guide.
Guo, T., X. Chen, Y. Wang, R. Chang, S. Pei, N. V. Chawla, O. Wiest, and X. Zhang, 2024, Large language model
based multi-agents: A survey of progress and challenges: arXiv preprint arXiv:2402.01680.
Hajizadeh, Y., 2019, Machine learning in oil and gas; a SWOT analysis approach: Journal of Petroleum Science and
Engineering, 176, 661-663.
Hassan, Q., P. Viktor, T. J. Al-Musawi, B. M. Ali, S. Algburi, H. M. Alzoubi, A. K. Al-Jiboory, A. Z. Sameen, H. M.
Salman, and M. Jaszczur, 2024, The renewable energy role in the global energy Transformations: Renewable
Energy Focus, 48, 100545.
He, K., G. Gkioxari, P. Dollár, and R. Girshick. 2017, Mask r-cnn. Paper read at Proceedings of the IEEE international
conference on computer vision.
Hong, S., X. Zheng, J. Chen, Y. Cheng, J. Wang, C. Zhang, Z. Wang, S. K. S. Yau, Z. Lin, and L. Zhou, 2023, Metagpt:
Meta programming for multi-agent collaborative framework: arXiv preprint arXiv:2308.00352.
Hoque, M., 2023, Understanding Artificial General Intelligence: The Frontier of AI.
https://medium.com/@minh.hoque/understanding-artificial-general-intelligence-the-frontier-of-ai-
de58228dbdde.
Hu, E. J., Y. Shen, P. Wallis, Z. Allen-Zhu, Y. Li, S. Wang, L. Wang, and W. Chen, 2021, Lora: Low-rank adaptation of
large language models: arXiv preprint arXiv:2106.09685.
Huang, S., W. yue, L. jun, W. fei, and Y. xin, 2024, Automatic Extraction of Standard Multimodal Knowledge for the
Petroleum Field: Highlights in Science, Engineering and Technology, 90, 209-217.
http://dx.doi.org/10.54097/e2xzf682.
Hussein, A., 2023, Chapter 1 - Oil and Gas Production Operations and Production Fluids, in A. Hussein, ed., Essentials
of Flow Assurance Solids in Oil and Gas Operations: Gulf Professional Publishing, 1-52.
Islam, M., and M. Khan, 2013, The petroleum engineering handbook: sustainable operations: Elsevier.
Jacobs, T., 2024, As Hype Fades, LLMs Gaining Acceptance in Upstream as New Age Research and Coding Tool:
Journal of petroleum technology, 76, no. 03, 34-41.
Jeon, W., G. Ko, J. Lee, H. Lee, D. Ha, and W. W. Ro, 2021, Deep learning with GPUs, Advances in Computers:
Elsevier, 167-215.
Jiang, W., X. Wang, and S. Zhang, 2023, Integrating multi-modal data into AFSA-LSTM model for real-time oil
production prediction: Energy, 279, 127935. http://dx.doi.org/https://doi.org/10.1016/j.energy.2023.127935.
Kanade, V., 2022, Narrow AI vs. General AI vs. Super AI: Key Comparisons.
https://www.spiceworks.com/tech/artificial-intelligence/articles/narrow-general-super-ai-difference/.
Khan, A. A., A. A. Laghari, and S. A. Awan, 2021, Machine learning in computer vision: a review: EAI Endorsed
Transactions on Scalable Information Systems, 8, no. 32, e4-e4.
Khan, S., M. Naseer, M. Hayat, S. W. Zamir, F. S. Khan, and M. Shah, 2022, Transformers in vision: A survey: ACM
computing surveys (CSUR), 54, no. 10s, 1-41.
Kingma, D., T. Salimans, B. Poole, and J. Ho, 2021, Variational diffusion models: Advances in neural information
processing systems, 34, 21696-21707.
Kirillov, A., E. Mintun, N. Ravi, H. Mao, C. Rolland, L. Gustafson, T. Xiao, S. Whitehead, A. C. Berg, and W.-Y. Lo.
2023, Segment anything. Paper read at Proceedings of the IEEE/CVF International Conference on Computer
Vision.
Koc, O., 2023, Artificial General Intelligence vs Narrow AI. https://www.linkedin.com/pulse/artificial-general-
intelligence-onur-koc/.
Koroteev, M., 2021, BERT: a review of applications in natural language processing and understanding: arXiv preprint
arXiv:2103.11943.
Latif, E., G. Mai, M. Nyaaba, X. Wu, N. Liu, G. Lu, S. Li, T. Liu, and X. Zhai, 2023, Artificial general intelligence
18
 (AGI) for education: arXiv preprint arXiv:2304.12479.
Li, J. X., M. Tsang, R. Zhong, J. Esterle, C. Pirona, M. Rajabi, and Z. Chen, 2023a, Automatic coal mine roof rating
calculation using machine learning: International Journal of Coal Geology, 274, 104292.
http://dx.doi.org/https://doi.org/10.1016/j.coal.2023.104292.
Li, W., Y. Ma, and D. S. R. Alerigi. 2024, Automated Wellhead Monitoring Using Deep Learning from Multimodal
Imaging. Paper read at International Petroleum Technology Conference.
Li, X., L. Zhang, Z. Wu, Z. Liu, L. Zhao, Y. Yuan, J. Liu, G. Li, D. Zhu, and P. Yan, 2023b, Artificial General
Intelligence for Medical Imaging: arXiv preprint arXiv:2306.05480.
Lin, T., Y. Wang, X. Liu, and X. Qiu, 2022, A survey of transformers: AI open, 3, 111-132.
Liu, C., Z. Liu, J. Holmes, L. Zhang, L. Zhang, Y. Ding, P. Shu, Z. Wu, H. Dai, and Y. Li, 2023a, Artificial General
Intelligence for Radiation Oncology: arXiv preprint arXiv:2309.02590.
Liu, X., H. Yu, H. Zhang, Y. Xu, X. Lei, H. Lai, Y. Gu, H. Ding, K. Men, and K. Yang, 2023b, Agentbench: Evaluating
llms as agents: arXiv preprint arXiv:2308.03688.
Lu, G., S. Li, G. Mai, J. Sun, D. Zhu, L. Chai, H. Sun, X. Wang, H. Dai, and N. Liu, 2023, AGI for Agriculture: arXiv
preprint arXiv:2304.06136.
Ma, Z., X. He, S. Sun, B. Yan, H. Kwak, and J. Gao, 2023, Zero-Shot Digital Rock Image Segmentation with a Fine-
Tuned Segment Anything Model: arXiv preprint arXiv:2311.10865.
Marlot, M., D. N. Srivastava, F. K. Wong, and M. X. Lee. 2023, Unsupervised Multitask Learning for Oil and Gas
Language Models with Limited Resources. Paper read at Abu Dhabi International Petroleum Exhibition and
Conference.
Meta, A., 2023, Introducing LLaMA: A foundational, 65-billion-parameter large language model: Meta AI.
Mishra, S., D. Khashabi, C. Baral, and H. Hajishirzi, 2021, Cross-task generalization via natural language
crowdsourcing instructions: arXiv preprint arXiv:2104.08773.
Muhammed, N. S., M. B. Haq, D. A. Al Shehri, A. Al-Ahmed, M. M. Rahman, E. Zaman, and S. Iglauer, 2023,
Hydrogen storage in depleted gas reservoirs: A comprehensive review: Fuel, 337, 127032.
Obaid, O. I., 2023, From machine learning to artificial general intelligence: A roadmap and implications:
Mesopotamian Journal of Big Data, 2023, 81-91.
Offert, F., and T. Phan, 2022, A sign that spells: DALL-E 2, invisual images and the racial politics of feature space:
arXiv preprint arXiv:2211.06323.
Ogundare, O., S. Madasu, and N. Wiggins. 2023, Industrial Engineering with Large Language Models: A Case Study
of ChatGPT's Performance on Oil & Gas Problems. Paper read at 2023 11th International Conference on
Control, Mechatronics and Automation (ICCMA), 1-3 Nov. 2023.
OpenAI, 2024, Creating video from text. https://openai.com/sora.
Oroojlooy, A., and D. Hajinezhad, 2023, A review of cooperative multi-agent deep reinforcement learning: Applied
Intelligence, 53, no. 11, 13677-13722.
Orr, J., and A. Dutta, 2023, Multi-agent deep reinforcement learning for multi-robot applications: A survey: Sensors,
23, no. 7, 3625.
Pal, S., K. Kumari, S. Kadam, and A. Saha, 2023, The ai revolution: IARA Publication.
Pandey, Y. N., A. Rastogi, S. Kainkaryam, S. Bhattacharya, and L. Saputelli, 2020, Machine learning in the oil and gas
industry: Mach Learning in Oil Gas Industry.
Park, J. S., J. O'Brien, C. J. Cai, M. R. Morris, P. Liang, and M. S. Bernstein. 2023, Generative agents: Interactive
simulacra of human behavior. Paper read at Proceedings of the 36th Annual ACM Symposium on User
Interface Software and Technology.
Paroha, A. D., and A. Chotrani, 2024, A Comparative Analysis of TimeGPT and Time-LLM in Predicting ESP
Maintenance Needs in the Oil and Gas Sector: International Journal of Computer Applications, 975, 8887.
Prestidge, K. L., 2022, Digital Transformation in the Oil and Gas Industry: Challenges and Potential Solutions,
Massachusetts Institute of Technology.
19
Rahate, A., S. Mandaokar, P. Chandel, R. Walambe, S. Ramanna, and K. Kotecha, 2023, Employing multimodal co-
learning to evaluate the robustness of sensor fusion for industry 5.0 tasks: Soft Computing, 27, no. 7, 4139-
4155.
Ruan, J., Y. Chen, B. Zhang, Z. Xu, T. Bao, G. Du, S. Shi, H. Mao, X. Zeng, and R. Zhao, 2023, Tptu: Task planning
and tool usage of large language model-based ai agents: arXiv preprint arXiv:2308.03427.
Shazeer, N., A. Mirhoseini, K. Maziarz, A. Davis, Q. Le, G. Hinton, and J. Dean, 2017, Outrageously large neural
networks: The sparsely-gated mixture-of-experts layer: arXiv preprint arXiv:1701.06538.
Sircar, A., K. Yadav, K. Rayavarapu, N. Bist, and H. Oza, 2021, Application of machine learning and artificial
intelligence in oil and gas industry: Petroleum Research, 6, no. 4, 379-391.
Speight, J. G., 2014, Handbook of offshore oil and gas operations: Elsevier.
Su, R., Q. Zhao, T. Zheng, G. Han, J. Jiang, and J. Hu, 2023, A Framework for RQD Calculation Based on Deep
Learning: Mining, Metallurgy & Exploration, 40, no. 5, 1567-1583.
Taori, R., I. Gulrajani, T. Zhang, Y. Dubois, X. Li, C. Guestrin, P. Liang, and T. B. Hashimoto. 2023, Stanford alpaca:
An instruction-following llama model.
Tariq, Z., M. S. Aljawad, A. Hasan, M. Murtaza, E. Mohammed, A. El-Husseiny, S. A. Alarifi, M. Mahmoud, and A.
Abdulraheem, 2021, A systematic review of data science and machine learning applications to the oil and gas
industry: Journal of Petroleum Exploration and Production Technology, 1-36.
Tarkowski, R., 2019, Underground hydrogen storage: Characteristics and prospects: Renewable and Sustainable
Energy Reviews, 105, 86-94.
Tatjana, P., 2017, Petroleum Extraction Engineering, in Z. Mansoor, ed., Recent Insights in Petroleum Science and
Engineering: IntechOpen, Ch. 2.
Tearpock, D. J., and R. E. Bischke, 2002, Applied subsurface geological mapping with structural methods: Pearson
Education.
Vaswani, A., N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, 2017, Attention
is all you need: Advances in neural information processing systems, 30.
VK, A., 2022, What Are the Types of Artificial Intelligence: Narrow, General, and Super AI Explained.
https://www.spiceworks.com/tech/artificial-intelligence/articles/types-of-ai/.
Waheed, U. 2023, Can ChatGPT write a review paper on full-waveform inversion? Paper read at 84th EAGE Annual
Conference & Exhibition.
Wang, H.-n., N. Liu, Y.-y. Zhang, D.-w. Feng, F. Huang, D.-s. Li, and Y.-m. Zhang, 2020, Deep reinforcement learning:
a survey: Frontiers of Information Technology & Electronic Engineering, 21, no. 12, 1726-1744.
Wang, L., C. Ma, X. Feng, Z. Zhang, H. Yang, J. Zhang, Z. Chen, J. Tang, X. Chen, and Y. Lin, 2023, A survey on large
language model based autonomous agents: arXiv preprint arXiv:2308.11432.
Wang, L., C. Ma, X. Feng, Z. Zhang, H. Yang, J. Zhang, Z. Chen, J. Tang, X. Chen, and Y. Lin, 2024, A survey on large
language model based autonomous agents: Frontiers of Computer Science, 18, no. 6, 1-26.
Waqar, A., I. Othman, N. Shafiq, and M. S. Mansoor, 2023, Applications of AI in oil and gas projects towards
sustainable development: a systematic literature review: Artificial Intelligence Review, 1-28.
Weijermars, R., U. bin Waheed, and K. Suleymanli, 2023, Will ChatGPT and Related AI-tools Alter the Future of the
Geosciences and Petroleum Engineering?: First Break, 41, no. 6, 53-61.
Wellmann, F., and G. Caumon, 2018, 3-D Structural geological models: Concepts, methods, and uncertainties,
Advances in geophysics: Elsevier, 1-121.
Wu, Q., G. Bansal, J. Zhang, Y. Wu, S. Zhang, E. Zhu, B. Li, L. Jiang, X. Zhang, and C. Wang, 2023a, Autogen:
Enabling next-gen llm applications via multi-agent conversation framework: arXiv preprint arXiv:2308.08155.
Wu, T., S. He, J. Liu, S. Sun, K. Liu, Q.-L. Han, and Y. Tang, 2023b, A brief overview of ChatGPT: The history, status
quo and potential future development: IEEE/CAA Journal of Automatica Sinica, 10, no. 5, 1122-1136.
XAI, 2024, Open Release of Grok-1. https://x.ai/blog/grok-os.
Xi, Z., W. Chen, X. Guo, W. He, Y. Ding, B. Hong, M. Zhang, J. Wang, S. Jin, and E. Zhou, 2023, The rise and potential
20
 of large language model based agents: A survey: arXiv preprint arXiv:2309.07864.
Youvan, D. C., 2024, Beyond Human Comprehension: AGI-Driven Advances in Information Geometry Using
Quantum Computing.
Zhao, W. X., K. Zhou, J. Li, T. Tang, X. Wang, Y. Hou, Y. Min, B. Zhang, J. Zhang, and Z. Dong, 2023a, A survey of
large language models: arXiv preprint arXiv:2303.18223.
Zhao, Y., A. Gu, R. Varma, L. Luo, C.-C. Huang, M. Xu, L. Wright, H. Shojanazeri, M. Ott, and S. Shleifer, 2023b,
Pytorch fsdp: experiences on scaling fully sharded data parallel: arXiv preprint arXiv:2304.11277.
Zhong, R., C. Salehi, and R. Johnson, 2022, Machine learning for drilling applications: A review: Journal of Natural
Gas Science and Engineering, 108, 104807. http://dx.doi.org/https://doi.org/10.1016/j.jngse.2022.104807.
Zohuri, B., 2023, Navigating the Global Energy Landscape Balancing Growth, Demand, and Sustainability: J Mat Sci
Apl Eng 2 (4), 01, 7.

21