In this article, Gaurav Agrawal, a member of the SPE Research and Development Technical Section (RDTS) explores the ever-expanding topic of machine learning (ML) and artificial intelligence (AI) with Zikri Bayraktar, SPE, from SLB’s Software Technology and Innovation Center (STIC) in Menlo Park, California.
This series highlights innovative ideas and analysis shaping the future of energy, with a focus on emerging technologies and their roadmaps, potential, and impact. With these conversations, we hope to inspire dialogue and accelerate progress across new energy frontiers.
Zikri Bayraktar is a senior machine learning engineer at the SLB Software Technology and Innovation Center (STIC). Before joining STIC, he spent 11 years at the Schlumberger-Doll Research Center as the AI research lead, where he led projects focused on automated reservoir steering, geology, carbon capture, and intelligent automation. Earlier in his career, he worked in IBM’s semiconductor research and development division. Bayraktar holds a PhD in electrical engineering and computational science from Penn State University and an MSc in management from the University of Illinois at Urbana-Champaign. He has coauthored 14 journal articles, 42 conference papers, a book chapter, and holds six patents. He is a senior member of IEEE and a member of SPE and SPWLA.
RDTS: AI and ML are often used interchangeably. Do they overlap?
Bayraktar: Though they overlap to a certain extent, they are in fact distinct. AI broadly focuses on systems that emulate the human decision-making process to solve problems, including rule-based systems, data-based algorithms in ML, robotics, and others.
ML can be considered a subset of AI, as algorithms that can learn from data, find patterns, improve outcomes, or automate certain tasks without explicit instructions. Both fields extensively utilize available data, may consume large computational resources, and may still produce stochastic outcomes at the end.
Three papers published in the 1950s were significant: “Turing Test,” “Three Laws of Robotics,” and “Perceptron” established the roots of neural networks (NN) from biological abstractions. A decade later, Perceptron trained with the backpropagation provided the foundations for the transformers that gave us the translators, chatbots, and AI/ML solutions that we enjoy today!
RDTS: Could you share examples of AI/ML-enabled solutions that were not possible some years back?
Bayraktar: I live in the San Francisco Bay area, where there are an increasing number of self-driving cars. Stepping into one of these cars and not seeing a driver is unnerving at first. But after a few minutes of smooth driving in complex urban traffic, you ease up and marvel at the technology. Partly driven by advanced ML algorithms and AI systems, which can seamlessly fuse various sensors and make decisions, self-driving cars are no longer a fantasy.
Similarly, personal assistants that can interact with human voice, cameras, and text are nowadays part of our daily lives: They can control home appliances, turn utilities on and off, warn about security issues, and even feed our pets. They are driven by high-accuracy sensor data, sophisticated large language models (LLMs), and AI agents.
RDTS: What level of computer power is needed to run AI/ML? Can we use AI/ML tools on our personal computers?
Bayraktar: It depends. Computational resources vary depending on the amount of available data, the AI model type, the number of ML model parameters, and how long a user is willing to wait to obtain good results.
Once the data is secured and its privacy ensured, depending on the architecture, ML models can be trained on laptops with central processing units (CPUs) or with graphics processing units (GPUs), desktop workstations, or clusters with hardware accelerators such as GPUs or tensor processing units (TPUs).
If the goal is inference using a trained model, a distilled or quantized AI/ML model, such as LLMs, can be deployed on personal computers or even mobile phones. Properly quantized AI/ML models require fewer hardware resources (RAM, GPU, etc.) without sacrificing performance much.
The largest models of the scale of ChatGPT or Gemini are usually run on shared servers with high-end GPUs, where speed of execution is important to the user. That said, we have also deployed tiny ML models in real-time scenarios that can fit in the small memory size of the field tools and even work with field-programmable gate arrays (FPGAs) or application-specific integrated circuits (ASICs).
RDTS: How can the AI/ML tools be accessed?
Bayraktar: Today, almost all cloud service providers and tech companies have their own versions of language and vision foundation models. The cost of accessing models both through programmable application programming interfaces (APIs) and user-friendly graphical user interfaces (GUIs) is dropping fast. These models run on cloud servers, and any data provided by the user is stored on third-party servers.
Open-source AI communities have come a long way. Microsoft’s Phi series and Meta’s Llama series are free versions of their sophisticated ML models. This makes it much easier for local deployment on private data and fosters greater adoption of the AI/ML technology.
Low-code or no-code auto ML tools are particularly useful for AI/ML newcomers. They offer a variety of algorithms where users bring in their data. The tool makes recommendations on how to clean and format data, select efficient models, and train them. It also helps analyze the results. I found these tools very helpful for training new models on limited datasets for quick prototyping. These software packages can be accessed on third-party servers or downloaded to personal computers.
RDTS: Where can we use AI/ML in oil and gas applications?
Bayraktar: Current LLMs are extremely versatile. We are quickly getting used to AI voice agents creating meeting summaries or digesting complex financial documents or contracts to extract correct information, and automating financial planning and budgeting applications.
In oil and gas, AI/ML applications help with the classification of log signals, quality checking of measured logs, borehole image and object classification, seismic data processing and automation, surrogate models for complex physics models, ML models for chemical synthesis or material screening, process control, automation, and many more.
One specific branch that I work on is the infusion or hybridization of physics models with ML-based models to capture more complex behavior than was possible before.
RDTS: You have published using AI/ML in log analysis and geologic modeling. Could you please share your experiences and the results?
Bayraktar: In the past decade, I developed, architected, and built various AI/ML models. In my earlier work, I built surrogate NN models that were orders of magnitude faster than the existing electromagnetics solvers. They were utilized in inversion problems. After this success, I started building NNs that input raw tool signals and generate interpretations of oil-based mud microresistivity tools, which were intended as inversion workflow replacements.
With these NNs, we achieved orders of magnitude speedup while matching the quality of inversion-based interpretation. We then branched into the classification of borehole images via convolutional neural networks (CNNs) based on the understanding of sedimentary geometry. Since then, we have published many more projects in material screening, automation of tasks, and design of language models that can understand our domain.
Through these studies, the main challenge has always been accessing good-quality data. Once the data hurdle is resolved, designing the ML project that solves a technical or business problem requires collaboration between the ML engineer and the subject matter experts, a well-defined success criterion, and a well-planned deployment strategy. At any stage of the workflow, a wrong decision can affect either the viability of the model or the deployment.
RDTS: What are the challenges in gathering good and sufficiently large training data?
Bayraktar: In our field, data is justifiably guarded. Accessing it to train ML models is a big challenge. If the amount of data is an issue, there are ML architectures with a low number of parameters that can provide good results. However, if the data lacks diversity, then the model will be biased and will not generalize. In such scenarios, local or basin-based models can serve the needs. There is no magic number for data size; one can always adapt the best model architecture, tune parameters, augment data, and place guardrails with known physics.
If a project or ML approach is new, derisking the approach with publicly available datasets makes sense. I stay cautious and make sure data licensing is handled prior to taking on the challenge. For proprietary data, one must go through the proper steps to safeguard data residency, data privacy, integrate siloed datasets efficiently, establish data governance for the long run, and ensure data labeling is globally consistent.
RDTS: Where do we need to collaborate to accelerate the development of AI/ML solutions for oil and gas?
Bayraktar: When it comes to successful AI/ML adaptation, our industry can collaborate with the cloud industry, academic institutions, government agencies, and open-source communities, as well as professional societies like SPE.
Cloud tech companies have the resources to support sharing open-sourced datasets for academic and research use. Data science companies can adapt their tools to the needs of our industry. Software companies focus on the industry-standard data models and make it easier to deal with large volumes of data.
Universities can encourage and teach how to leverage the intricate domain-specific knowledge in AI projects, and government agencies can make it easier to share knowledge and data, potentially around the globe.
Professional organizations such as SPE can facilitate this knowledge sharing, advocate policies to accelerate AI collaboration, provide low-cost training and certification programs, and be the collaborative platform for everyone from students to oil and gas veterans to contribute to next-generation AI/ML developments.
RDTS: What is the role of the open-source community in future developments of AI/ML?
Bayraktar: I am a big supporter of open-source, and I truly believe that it increases both the pace of development and the usage of AI/ML methods. At various times, I took inspiration from AI methods deployed in other domains, enjoyed open-source examples to ease my learning curve, and adapted solutions that were seemingly irrelevant to oil and gas at first glance. Studies have shown that open-source software practices can contribute immensely to the economy, create fierce competition, and allow transparency for safe and responsible models.
One of the successful open-source platforms in the AI/ML space is Hugging Face, which I have been following since its founding days.
They not only created a platform for sharing models but also unified model development to a certain extent, built open communities around various topics, and shared every aspect of ML development, starting from data to deploying models online to building free educational resources around some of the cutting-edge topics in the field. Their combined approach to open-source and communities made a significant contribution to various developments today, which is commendable.
RDTS: What are the training resources and career pathways to become an AI/ML expert in oil and gas?
Bayraktar: If you are a trained scientist or engineer, there is a good chance that you already have taken fundamental math and programming courses to branch into ML. Currently, there are various open-source courses from top universities on YouTube and their respective course websites.
I highly recommend matching the resources to their project at hand and learn accordingly to immediately apply learnings to a real-world problem. I found various low-cost online courses with practical coding assignments very useful, both in terms of understanding the underpinning theory as well as for the practical experience aspects of training a model based on the data at hand.
Similarly, SPE has various sources available for learning and practicing AI/ML. My first project with ML, for which we obtained a patent, was using simple NNs, but we had to build the workflow from the ground up based on domain expertise. If someone has an AI/ML degree, they can partner with domain experts to apply their skills to solve complex scientific and engineering problems as well.
RDTS: What new developments in AI/ML are you personally looking forward to?
Bayraktar: Generative AI agents are a new and high-impact development that makes using AI easier. Using the available tools and without explicit instructions, it can proactively reason to achieve the user-assigned goals. These specialized GenAI models can access an API, a database or a web service, a custom proprietary function, simple Internet of Things (IoT) devices, or large systems. I believe we are only starting, and this will expand rapidly.
Additionally, there are several AI research topics that I’d like to see flourish. One that is pertinent to oil and gas is related to the multimodality of models. We humans learn and process data that is in various modalities.
AI can do a better job utilizing such datasets. While LLMs opened the door to new avenues, I think that text should go along with image and sound as well to achieve even more complex tasks. In oil and gas, we have various data modalities, and we can benefit immensely if AI research focuses on handling a variety of unstructured data modalities.
Outside our field, I am looking forward to seeing further achievements in AI weather prediction models that can save lives and reduce property damage. In this front, Fourier neural operators, diffusion models, and graph neural networks show great promise both at local and global scales.
Read the first article in this series, which discussed a closely related topic in a Q&A with Madhava Syamlal, CEO and founder of QubitSolve, Quantum Computing—Are We Ready?
