A physics-constrained deep learning model for simulating multiphase flow in 3D heterogeneous porous media

Physics-based simulators for multiphase flow in porous media emulate nonlinear processes with coupled physics, and usually require extensive computational resources software development, maintenance simulation execution. As a result, huge demand exists fast modeling of wide range subsurface applications including geological CO2 sequestration, hydrocarbon recovery geothermal energy extraction. In this work, an efficient physics-constrained deep learning model is developed solving 3-Dimensional (3D) heterogeneous media. The fully leverages the spatial topology predictive capability convolutional neural networks, specifically U-Net successive contracting expansive steps, continuity-based smoother to predict responses that need continuity. Furthermore, transient regions are penalized steer training process such can accurately capture these regions. takes inputs properties media, fluid well controls, predicts temporal-spatial evolution state variables (pressure saturation). While maintaining continuity flow, 3D domain decomposed into 2D images reducing cost, decomposition results increased number data samples better efficiency. Additionally, surrogate separately constructed as postprocessor calculate rate based on predictions from model. We use example injection saline aquifers, apply trained physics-based emulates physics process. performs prediction speedup ? 1400 times compared simulations, average temporal errors predicted pressure saturation plumes 0.27% 0.099% respectively. water production efficiently by rate, mean error less than 5%. Therefore, its unique scheme cope fidelity become computationally demanding inverse problems or other processes.