SimGraph: A Scalable Physics-Informed Graph Neural Network-Based Smart Proxy for Reservoir Simulation Workflows

In this paper, we introduce SimGraph, a graph neural network-based proxy for the dynamic simulation of subsurface reservoirs. SimGraph decomposes reservoir geometry into sectors and utilizes graphics processing units (GPUs) for parallel processing. This sectorization approach allows the model to extrapolate to new reservoir models with minimal loss of accuracy. A key feature of SimGraph is the consideration of physics constraints while formulating the loss term during training, which ensures that the simulated flow adheres to these constraints. This results in physically consistent outputs. Sim-Graph supports both forward and inverse modeling workflows and is particularly effective in Assisted History Matching (AHM) exercises.We present detailed hardware scaling studies of SimGraph, demonstrating its performance on the SPE10 reservoir model, which consists of approximately 1.2 million cells. Our results are benchmarked against other legacy mesh-based solvers, highlighting SimGraph’s scalability and accuracy. This work demonstrates SimGraph’s potential as a robust tool for reservoir simulation in high-performance computing environments. The ability to efficiently utilize multiple GPUs and maintain accuracy across different geological models highlights its practical value.