- Institute of energy technology, Department of Reservior technology, Kjeller, Oslo, Norway (hongliang.wang@ife.no)
Simulating subsurface geological processes, such as compaction-driven fluid flow and rock deformation, is essential for understanding natural phenomena and addressing challenges in energy production and resource extraction. Traditional numerical methods, while effective, are computationally expensive and struggle to efficiently model large-scale dynamical problem in subsurface systems. This creates a bottleneck for large-scale simulations and real-time decision-making. Recent advances in machine learning (ML) offer promising solutions to enhance simulation efficiency. Neural operators, which learn mappings between function spaces, provide a flexible, scalable approach to modeling complex systems. Unlike traditional methods, neural operators can generalize across varying inputs and geometries, offering a more efficient and versatile alternative. This study explores the potential of advanced machine learning techniques, specifically Fourier Neural Operators (FNO) and Physics-Informed Neural Operators (PINO), to model two critical subsurface geomechanical processes: compaction-driven fluid flow and elastic stress analysis for tunnelling.
For the first case, numerical simulations were conducted to generate a dataset of up to ~10,000 samples, derived from ~1,000 different initial porosity conditions represented by randomly generated polygons. The FNO model was trained using Nvidia A100 GPUs (80G). Training loss decreased rapidly during early epochs and stabilized below 0.02 after approximately 50,000 epochs. Models trained with larger datasets (e.g., 9,753 samples) demonstrated significantly improved validation performance, achieving a validation loss of ~0.06. In contrast, models trained on smaller datasets exhibited overfitting, with validation losses exceeding 0.3. The trends in validation loss, evaluated using 60 test cases with elliptical initial conditions excluded from the training data, underscored the importance of dataset size in enhancing model generalization for machine learning-based geological simulations. The validation results demonstrated high predictive accuracy, with maximum errors below 10%. Models trained on larger datasets achieved superior performance, particularly for cases with sharper structural features. These findings highlight the capability of FNO models to effectively generalize and reproduce the dynamics of complex fluid flow in subsurface environments. The second case focuses on elastic stress analysis for tunneling, where stresses and deformations around underground excavations are critical to ensuring structural stability. Preliminary efforts have been directed toward generating numerical datasets to train FNO and PINO models, with the goal of capturing stress distribution and deformation patterns under diverse geological and engineering conditions. While results are still emerging, early indications suggest that PINO may provide additional advantages by incorporating physical laws directly into the training process, potentially reducing the amount of data required and improving computational efficiency.
This work demonstrates the transformative potential of neural operators in addressing computational challenges associated with subsurface geomechanical modeling. By combining the flexibility of data-driven methods with the robustness of physics-informed approaches, FNO and PINO offer scalable and efficient alternatives to traditional numerical methods.
How to cite: Wang, L. H. and Yarushina, V.: Harnessing Neural Network and Operators to Simulate Subsurface Geomechanical Processes , EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-13428, https://doi.org/10.5194/egusphere-egu25-13428, 2025.