A Graph Neural Network-Based Approach for Complex Reservoirs Simulation Surrogate Modelling
Résumé
Reservoir modeling and simulation are crucial in geoscience applications such as predicting oil production or CO2 plume migration under geological uncertainty. ++This study presents an innovative approach that uses graph neural networks (GNN) to improve the accuracy of forecasting key reservoir properties, pressure, and saturation. ++The developed model is trained on a few hundred simulation runs with different geological stochastic realizations and it shows a remarkable ability to predict reservoir dynamics for new unseen realizations with a speed factor of more than a thousand times. ++This ability is crucial for both improving hydrocarbon extraction by history matching or field development optimization and for carbon sequestration risk assessment. + +Different from previous methods, this study contributes further to the field by combining neural networks with graph-based learning, which has not been widely explored in previous works. ++It also combines the surrogate models obtained for the saturation and pressure fields in a novel way to mimic numerical methods approximation of multiphase flow equations. + +Several synthetic datasets from CCS (Carbon Capture and Storage) and oil reservoirs also used in previous works are used to test the robustness of the approach, particularly concerning the proposed architecture. ++The results show that the GNN model can provide more accurate predictions of complex reservoir dynamics compared to traditional convolutional neural networks. ++Important technical contributions of this research include the development of a GNN model specifically adapted to analyze spatio-temporal geological data by mixing both saturation and pressure predictions applicable in both oil production and CO2 sequestration applications. ++Additionally, the model has been adapted to complex unstructured and irregular reservoir meshes, which is a significant improvement over previous approach. + +This research presents an innovative approach for reservoir modeling, setting a benchmark in the field and offering significant implications for both hydrocarbon production and CO2 sequestration. ++By their impressive computational speed, deep learning surrogates can revolutionize the way reservoir modeling and simulation are performed, leading to more accurate and efficient predictions and ultimately better decision-making.