The energy transition demands smarter, more efficient ways to assess the subsurface, and artificial intelligence (AI) is transforming how geophysical data — particularly seismic and downhole logs — is analyzed and interpreted. Traditional methods often struggle to handle the sheer volume and complexity of geophysical datasets, leading to gaps in subsurface predictions. AI bridges this gap by automating feature detection, improving lithology classification, and enhancing the interpretation of rock properties in real time. For seismic data, AI-driven models can identify subtle geological features critical for optimizing geothermal reservoirs or CO₂ storage sites. In the case of downhole logs, AI algorithms offer faster, more accurate predictions of reservoir characteristics, even in complex carbonate or fractured systems. By combining geophysical data with machine learning, we can significantly reduce exploration risks, improve reservoir management, and ensure more reliable deployment of energy transition technologies across diverse geological settings.

Core samples provide geologists with high-resolution insights into the textures and fabrics of sedimentary rocks, which are essential for subsurface characterization and reservoir analysis. However, core recovery is often limited due to high costs or poor recovery rates during drilling operations, creating a significant gap in the geological data available for many wells. An alternative solution is to utilize high-resolution wireline imaging tools. These provide a continuous record of the subsurface, but at the cost of interpretability. Our group has focused on automating the interpretation of downhole logs using artifical intelligence.

An example of this is our work on Formation MicroScanner (FMS), which provides detailed resistivity-based images of borehole walls. Yet, interpreting FMS image logs requires specialized skills that not all geologists possess. To address this challenge, our team has developed a machine learning solution using Generative Adversarial Networks (GANs) to automatically generate realistic core images from FMS logs (Baharudding and John, submitted).

Examples of real (left) and generated (right) core images from our trained GANs. FMS data is used as a conditional input to generate the synthetic cores. Both images are from the same target interval, demonstrating the usefulness of our approach.

By training a series of image-to-image translation models, including unsupervised and supervised GANs, we were able to produce synthetic core images that closely match real core samples in terms of texture, detail, and geological accuracy. Among the models tested, the pix2pixHD model trained on concatenated FMS pad images provided the most realistic core images, with low Root Mean Square Error (RMSE), high Peak Signal to Noise Ratio (PSNR), and high Structural Similarity Index Method (SSIM) scores when compared to ground truth core images. This approach significantly improves the interpretation of sedimentary facies, as blind testing with geologists showed a dramatic improvement in classification accuracy from 14% on raw FMS logs to 73% on synthetic core images.

Our research highlights the potential of GANs to bridge the gap between limited core samples and the need for comprehensive geological facies classification. By augmenting traditional data acquisition with AI-driven synthetic core images, this work has the potential to transform subsurface characterization workflows, reduce reliance on physical core samples, and enable geologists to make more accurate interpretations even in data-scarce environments. As digital tools continue to evolve, this approach offers a cost-effective and scalable solution for improving geological insights and reservoir modeling in both exploration and production settings.

Another very important source of subsurface information is seismic data. Our research group is actively developing machine learning-based solutions to overcome long-standing challenges in seismic data acquisition and interpretation, which are essential for both hydrocarbon exploration and geological studies. One of the key issues in seismic data processing is dealing with sparse or incomplete seismic surveys, which can introduce interpretational biases and reduce the effectiveness of downstream applications such as seismic facies analysis. Traditional methods for interpolating missing traces rely heavily on expert-driven processes that assume geologic event separability, but these approaches often fail to capture the complex structures in the subsurface. To address these limitations, we developed a deep learning-based method using a Wiener-filter loss function that improves the accuracy of seismic trace reconstruction by focusing on signal structure and amplitude preservation rather than pixel-level accuracy (AlSalmi and John, submitted). This novel method outperforms traditional loss functions like L1 and L2, resulting in sharper geological features and better seismic image resolution. By improving the quality of reconstructed seismic data, this approach has the potential to reduce costs associated with reshooting seismic surveys and enhance exploration workflows.

Illustration of the SRT-Ai algorithm applied to a seismic line offshore Australia (Al Gharbi et al, in review)

We also currently work on automation of seismic stratigraphic interpretation using deep learning models. The process of seismic stratigraphy, introduced in the 1970s, revolutionized the field by enabling geologists to identify seismic reflection terminations and infer past base level changes and sedimentary processes. However, this method remains labor-intensive, requiring visual inspection of thousands of seismic inlines and crosslines, and is highly dependent on interpreter expertise. To address this, we developed a Convolutional Neural Network (CNN) model named SRT-Ai (Seismic Reflection Terminations Attribute), which was trained on 160,000 synthetic seismic models representing different types of seismic reflection terminations such as truncation, onlap, toplap, and downlap (Al Gharbi et al, submitted). SRT-Ai achieved 99.9% accuracy and precision on synthetic test data and demonstrated strong generalization potential when applied to real-world seismic datasets from northwest Australia, achieving 91% accuracy and 96% precision. This tool provides automated, scalable identification of seismic boundaries, significantly reducing interpretation time and increasing the consistency of early-stage seismic reconnaissance.

Both projects highlight the potential for machine learning to transform seismic workflows, by offering cost-effective solutions to data reconstruction and automated stratigraphic interpretation. These innovations will not only improve the efficiency of geological exploration but also help reduce uncertainty in seismic interpretations, ultimately leading to better-informed decisions in resource management and geohazard assessment.

Natural gas from North American shale reservoirs offers a more sustainable energy solution compared to imported liquefied natural gas (LNG) or crude oil from the Middle East. Locally sourced shale gas reduces the carbon footprint associated with transportation and is a cleaner-burning alternative to coal and oil, making it a key transitional fuel in reducing emissions. However, extracting gas efficiently from heterogeneous shale formations remains a challenge, requiring precise identification of productive wells to minimize environmental impact.

In a book chapter recently published, we show that conventional machine learning (or “statistical machine learning”) can be used to predict production in U.S. Gas Shale. This is important to move away from more polluting fossil fuels (such as petroleum), and to have a secure supply of energy.

Recent work within our research group has shown that AI-driven geophysical analysis can play a transformative role in optimizing shale reservoir production. By applying machine learning (ML) algorithms to well log data across seven major U.S. shale plays, we demonstrated that random forest (RF) classifiers outperform other methods in predicting well success. Our refined RF model achieved 97% accuracy in forecasting initial production rates, using features such as average play thickness, pore pressure, and total vertical depth (TVD). This research highlights how AI can help address subsurface variability, improve well targeting, and reduce exploratory drilling.

As the energy transition progresses, AI tools are becoming essential for smarter and more sustainable shale gas extraction. By integrating AI with local resource development, we can reduce reliance on higher-emission imports while ensuring that natural gas remains a key transitional energy source with lower environmental impacts than coal or oil.