EAGE Workshop on Enhancing Subsurface Practices using AI/ML

This paper introduces an end-to-end in-house development of an AI-assisted subsurface workflow designed for compartmentalized gas fields in the Gulf of Thailand, which faces the challenge of small reservoirs requiring a large number of wells and swift subsurface analysis. The objective is to streamline and automate key subsurface processes—subsurface interpretation (fault, horizon, prospect), well targeting, and production forecasting—to reduce exhaustive effort, improve consistency, and suggest optimal decision-making across the full field development lifecycle.

It blends domain expertise with advanced analytics, runs on HPC, and integrates with commercial software, offering a scalable solution for faster, smarter E&P decisions. The overall could streamline the process in GoT and provide benefit from improved investment decisions in all aspects, from seismic interpretation, well targeting, and production optimization.

The paper presents the application of machine learning (ML) techniques to a complex deep-water Sarawak dataset to accelerate the prediction of the Mid Miocene Unconformity (MMU) for Earth model building (EMB). Leveraging locally trained convolutional neural networks (CNNs) and transfer learning with sparse local labels, the workflow improves horizon tracking and significantly reduces manual interpretation effort in high-confidence areas. The study demonstrates the integration of regional ML models into EMB workflows in geologically complex basins, while highlighting the essential role of human expertise in low-confidence zones—redirecting effort toward higher-value activities such as data evaluation and advancing subsurface understanding critical to petroleum system analysis.

Surface-related multiples contaminate seismic data and degrade imaging quality. Traditional suppression methods suffer from limited ability or high computational demand, while supervised neural networks require clean training labels which are unavailable for real data. To address these issues, we propose a self-supervised learning framework using a two-stage training strategy with warm-up and iterative data refinement phases. Our method requires only single multi-dimensional convolution to generate synthetic multiples, eliminating dependency on clean labels or velocity models. The network progressively learns to suppress multiples while preserving primary reflections through epoch-based refinement cycles. Validation on real marine seismic data demonstrates effective multiple attenuation. Migration results confirm removal of spurious artifacts and enhanced subsurface imaging. This approach provides a practical, flexible solution for multiple suppression in real-world seismic processing.

The Barrow and Dampier sub-basins of the Northern Carnarvon Basin in Western Australia have emerged as promising regions for geological carbon dioxide (CO2) storage. Permeability, a key petrophysical property for assessing CO2 injectivity, is difficult to measure directly from well logs and typically requires limited core analysis or well testing. This study evaluates the effectiveness of the XGBoost machine learning algorithm in predicting permeability using well logs and interpreted petrophysical properties. A well-based data splitting strategy is adopted to prevent data leakage and to rigorously assess model generalisation across diverse geological settings. Although most training data originated from the Mungaroo formation, the model also performs well on other key CO2 storage formations, suggesting that the variability within the Mungaroo dataset, supplemented by additional data from CO2 storage formations, is broad enough to capture common features across them. Furthermore, integrating petrophysical interpretation results—such as effective porosity and mineral composition—with raw logs significantly enhances prediction accuracy. These findings demonstrate the potential of combining data-driven modelling with domain knowledge to enable reliable, regional-scale permeability prediction in support of CO2 storage formation screening.

In order to handle large volumes of legacy borehole petrophysics data, the authors propose Skippy, a subject matter expert-guided tool for LAS file harmonization. The tool comprises a set of modular steps to handle discrete components of the harmonization workflow, allowing adaptability and extensibility. Configuration of the tool is performed in plain text JSON files, suitable for non-coders.

The tool has been developed using the Geological Survey of Western Australia’s WAPIMS database, and has generated initial harmonised outputs of the log data within.

An open-source release model is planned, inviting contributions.

We present a Physics-Informed Denoising Diffusion Probabilistic Model (PIDDPM) for super-resolution and surrogate modelling of time-dependent physical systems. PIDDPM is conditioned on a coarse-resolution input at the current timestep and high-resolution ground truth from the two preceding timesteps, with the aim of reconstructing fine-scale solutions consistent with the underlying dynamics. PIDDPM acts as a surrogate, approximating the behaviour of nonlinear PDEs such as the Allen-Cahn equation without requiring full numerical simulation. Physics-based penalties are incorporated into the loss function to penalise lack of consistency with the governing equations and boundary conditions, ensuring that the generated outputs remain physically plausible. Our results demonstrate that PIDDPM significantly improves perceptual and physical accuracy compared to baseline DDPM, achieving higher PSNR (Peak Signal-to-Noise Ratio) and SSIM (Structural Similarity Index Measure) while reducing MSE (Mean Squared Error) and MSGE (Mean Squared Gradient Error). The model’s ability to learn temporal evolution and spatial refinement makes it a scalable and physically grounded alternative to traditional solvers. PIDDPM shows strong potential for resolution enhancement, interpolation and predictive modelling in subsurface workflows, offering a data-driven approach to accelerate simulations and support efficient decision-making in geoscientific domains.

Reliable reservoir characterization requires high-quality, consistent data, and as machine learning (ML) becomes more embedded in geoscience, robust preparation and curation are essential. This study presents an integrated workflow that combines ML-assisted log QC, reconstruction, and rock typing with automated static reservoir model assessment. Using algorithms such as Isolation Forest, stratigraphy-based clustering, and Random Forest, the approach corrects anomalies, fills missing intervals, and predicts rock properties in uncored wells, enabling more complete and reliable 3D geological models. Automated outputs, including isochore maps, zonal validations, and log-derived comparisons, improve accuracy, highlight inconsistencies, and assess the impact of new well data on model reliability. The workflow reduces manual effort, accelerates project timelines, and increases usable well data by leveraging records previously discarded for poor quality. By using ML with traditional geoscience analysis, this methodology enhances real-time decision-making, optimizes well placement, and improves hydrocarbon recovery, setting a new standard for data-driven reservoir evaluation.

Seismic reservoir characterization aims to integrate various geophysical datasets to predict better subsurface models. Usually, seismic inversion workflow integrates wells and seismic data together to predict elastic models of the subsurface. However, elastic models are further interpreted into reservoir properties using traditional multi-linear regression or neural networks thereby increasing turnaround time and reducing precision as these reservoir properties are predicted one by one.

The latest integration of rock physics with machine learning enables geoscientists to use convolutional neural network (CNN) in reservoir characterization. This allows for the simultaneous prediction of both elastic and reservoir properties directly from seismic gather data. Integrating rock physics into the workflow is crucial as it allows us to simulate various geological scenarios in the subsurface. By doing this, we can predict the corresponding elastic properties, which in turn allows for the effective use of deep learning in reservoir characterization.

This study presents an AI-driven approach to petrophysical interpretation for enhanced reservoir characterization, using subsurface data from the Indus Basin, Pakistan. Traditional petrophysical analysis methods often face challenges related to data quality, non-linearity, and complex reservoir heterogeneity. To overcome these limitations, this research integrates Artificial Intelligence (AI) and Machine Learning (ML) techniques to predict key reservoir properties including porosity, water saturation, Lithofacies and shale volume.

Well log and, where available, was preprocessed, normalized, and used to train and test various ML models. Model performance was evaluated using standard statistical metrics such as R², RMSE, and MAE to ensure reliability. The results demonstrated that AI-based models significantly improve prediction accuracy compared to traditional empirical methods, especially in complex lithological zones.

The case study from the Indus Basin highlights the applicability of ML algorithms in identifying sweet spots, improving reservoir quality mapping, and supporting better decision-making in exploration and development planning. This research contributes to the growing body of work focused on digital transformation in the energy sector and showcases the potential of AI/ML in modern petroleum engineering workflow.

In this study, we trained a cycle-consistent adversarial network to translate forward-modelled synthetic DAS images into more realistic versions. The CycleGAN generator preserves the structure and positions of the synthetic events and automatically applies dataset-specific characteristics. After training, the generator can synthesize any number of realistic samples from a pool of synthetic events and labels can be directly transferred. The data distributions of these datasets more closely resemble those of the target datasets, which can be used to counteract model generalization issues.