Applying Two-Grid Preconditioner for Subsurface Flow Simulation using Attention-enhanced Hybrid Network to Accelerate Multiscale Discretization in High-contrast Media
Researchers have developed a novel hybrid framework that integrates machine learning with multiscale numerical methods to efficiently solve Darcy equations in complex subsurface flow simulations. The approach uses an attention-enhanced neural network to predict multiscale basis functions, significantly accelerating the offline computation stage of the mixed generalized multiscale finite element method (mixed GMsFEM). This learning-based acceleration, combined with a two-grid preconditioned solver for the global system, maintains accuracy and stability even in heterogeneous media with high-contrast coefficients, outperforming existing learning-based methods. AI
IMPACT Accelerates complex simulations, potentially enabling higher-resolution subsurface analysis for fields like geology and reservoir engineering.