During the development of oil and gas reservoirs, only a small percentage of the subsurface is sampled because of the high costs associated with these projects. Subsurface modeling, adjusted through history matching, approximates the subsurface by interpolating between sparse measurements at the wellbore. Traditional geostatistics and multipliers are used to build these models.
However, these commonly used approaches are limited to statistics and do not include geological understanding of the subsurface. Solving a history-matching inverse problem can end up with a statistical solution that violates the conceptual understanding of the subsurface (e.g., increasing or decreasing permeability in a specific region of the reservoir to match reservoir response).
Generative artificial intelligence (GenAI) is proposed to solve these problems. These models can be trained to learn the spatial geological depositional information and generate high-resolution subsurface models from a low-dimensional random vector (latent space). Given a latent space, the model is trained to create subsurface models using training subsurface models simulated with advanced subsurface modeling techniques.
The trained model can only generate subsurface models consistent with training data (mostly similar), which bounds them within the geological conceptual model understood by geoscientists. The latent space parameterization has a lower dimension compared to the subsurface model. This parameterization makes the history matching exercise easier as the optimization algorithm in the automated history matching only changes a lower dimensional representation to bring the subsurface model closer to reality.
Research at The University of Texas at Austin focuses on solving two challenges of applying GenAI for subsurface modeling and history matching. The first challenge is that the application of GenAI for subsurface modeling is checked mainly by visual inspections, which creates a need for a robust checking protocol. The second challenge is that several GenAI models exist, and their performance compared to each other is poorly understood, which creates a need to compare their performance to set model selection criteria.
A Minimum Acceptance Checks Protocol
The proposed minimum acceptance checks protocol focuses on global univariate distribution, spatial continuity, local and global uncertainty model, local exactitude, and fluid-flow simulation. The checks are applied before and after conditioning the GenAI to measure attributes at well locations. The conditioning-related artifacts and discontinuity are also investigated.
The proposed protocol allows for quantitative checks for better application of subsurface modeling. It avoids subjective judgment and increases confidence in the GenAI model. It is recommended that if GenAI models fail these checks, their performance must be improved before utilizing them.
● Global Univariate Distribution: Evaluated by histogram reproduction using the overlap coefficient and Q-Q plot divergence area.
● Local and Global Uncertainty Models: Assessed using expected per-pixel standard deviation and image comparison in multidimensional scaling space with the coefficient overlap.
● Spatial Continuity: Checked by reproducing variograms and measuring the dispersion ratio.
● Local Data Exactitude: Calculated by assessing conditioning error at measured data locations.
● Dynamic Response: Evaluated by running flow simulations for various spatial continuity ranges.
GenAI for Subsurface Models Comparison
The GenAI model represents a constantly evolving area of research, with numerous models emerging regularly. However, three main models are generally proposed, including variational autoencoders (VAE), generative adversarial networks (GAN), and diffusion probabilistic models (DDPM). These models are applied for subsurface modeling, and the literature reports conflicting results, mostly relying on visual inspection or limited qualitative metrics (Fig. 1).
When checked against the proposed minimum acceptance criteria, the models show some inherent limitations. VAEs have the worst performance in terms of global univariate distribution, local and global uncertainty, spatial continuity, and dynamic flow response. GANs exhibit a robust performance against all criteria and show the best performance of all models. DDPM performs well against most criteria except for the local uncertainty model.
The selection of the models is task specific. Choosing the right model depends on the task at hand. However, applying GAN for subsurface modeling is generally superior to all other models (Table 1).
Conclusion
The application of GenAI for subsurface modeling is a promising technology. It provides a method to balance geological understanding and conditioning abilities. However, the models need to be checked before trusting the generated realizations. The proposed minimum acceptance protocols and model selections allow quantitative approaches to select and check the models before their application to ensure robust decision-making processes.
Future research involves the development of checking methods and the application of GenAI for solving history-matching problems.
For Further Reading
GenAI Minimum Acceptance Checks Static and Dynamic Model Checking of Conditioning Generative Artificial Intelligence Models for Subsurface Modeling by A. Merzoug, L. Liu, M.J. Prycz, The University of Texas at Austin.
Conditional Generative Adversarial Networks for Subsurface Modeling: How Good They Really Are? by A. Merzoug, M.J. Prycz, The University of Texas at Austin.
Ahmed Merzoug is a PhD student in the Department of Petroleum and Geosystems Engineering at The University of Texas at Austin.
Merzoug’s research focuses on applying generative artificial intelligence to subsurface modeling and data assimilation, specializing in optimization and proxy modeling for unconventional oil and gas systems. His work aims to enhance resource recovery, improve operational efficiency, and streamline computational workflows for subsurface energy systems.
In addition to his academic and research endeavors, he has made significant contributions to the field with more than 25 publications covering unconventional oil and gas optimization, subsurface modeling, and enhanced geothermal systems (EGS). His work on EGS has advanced the understanding and optimization of geothermal resource development, further demonstrating his commitment to innovation in energy sustainability.
Through his academic achievements, research output, and dedication to solving complex energy challenges, he continues to drive technological advancements in energy exploration and production.
He holds a BS and an MS in petroleum engineering from the University of Boumerdes and a master’s in engineering from the University of North Dakota.
Michael J. Pyrcz is a professor at The University of Texas at Austin, with joint appointments in the Cockrell School of Engineering and the Jackson School of Geosciences. He leads the DiReCT Consortium (Digital Reservoir Characterization Technology), which focuses on integrating geostatistics, machine learning, and data analytics into subsurface reservoir characterization to optimize resource management and improve decision-making under uncertainty.
Pyrcz’s research emphasizes enhancing spatial subsurface modeling through advanced data science methods, including geostatistics and machine learning, to improve resource extraction and minimize environmental impacts. He and his students develop workflows that integrate engineering and geoscience data, account for spatiotemporal aspects, address scaling issues, and quantify uncertainty to support optimal decisions.
A dedicated educator, Pyrcz shares his educational content online through “GeostatsGuy Lectures,” providing access to cutting-edge knowledge worldwide. With over 60 peer-reviewed publications and coauthorship of Geostatistical Reservoir Modeling, he has significantly influenced geostatistics and subsurface analytics.
His work has earned him honors such as the SPE Gulf Coast Section Regional Data Science and Engineering Analytics Award (2021) and recognition as a Distinguished Lecturer by IAMG (2024). His teaching, research, and outreach drive innovation in sustainable resource management and data science integration in geosciences.