Semester of Graduation
Spring 2025
Degree
Master of Science (MS)
Department
Petroleum Engineering
Document Type
Thesis
Abstract
Reservoir simulation is the state-of-the-art method for predicting the flow of petroleum reservoir fluids in porous media. It provides an accurate and unbiased prediction of the performance of petroleum reservoirs under different operating conditions. Despite its advantages, reservoir simulation is computationally expensive; with typical full-field simulation models running for several hours. This limitation is worsened when simulating reservoirs with several equations for each cell, such as multiphysics or compositional reservoir simulation. The goal of this research is to provide a fast and accurate spatiotemporal machine-learning model that incorporates discretized governing mass balance equations for training. To achieve this, we propose two spatiotemporal graph neural networks: a heterogeneous graph neural network with long short-term memory (HGNN-LSTM) and a heterogeneous graph neural network with a transformer (HGNN-Transformer). The HGNN-LSTM model captures the spatial component of a reservoir simulation input data using an HGNN and represents the time dependence using the LSTM model. In contrast, the HGNN-Transformer model adopts a factorized STGNN structure that combines an HGNN for spatial learning with a Transformer model for temporal sequence modeling. Both models incorporate the discretized governing mass balance equations through a custom loss function, which enables them to generate predictions that honor these conservation laws.
To assess the effectiveness of these models before and after the governing physics, we compared them to corresponding predictions from an open-source reservoir simulation model. Before incorporating physics, the HGNN-LSTM model achieved lower training and validation losses than the HGNN-Transformer. After incorporating physics, we applied a design of experiments (DOE) approach to optimize weight configurations for pressure data loss, rate data loss, and residual loss. The results indicate that the weight for the physics component of the loss function should be higher towards the end of the training epochs than in the beginning. Comparing the best HGNN-LSTM results with and without physics shows that incorporating physics yields a 42% reduction in the residual or mass balance error. In contrast, incorporating physics into the best HGNN-Transformer model yielded a 13% reduction in the residual.
Date
12-23-2024
Recommended Citation
Abdullah, Ahmed A.M.A., "Physics-Informed Heterogeneous Spatiotemporal Graph Neural Network for Reservoir Simulation" (2024). LSU Master's Theses. 6082.
https://repository.lsu.edu/gradschool_theses/6082
Committee Chair
Olorode, Olufemi