A Physics‐Informed Deep Learning Framework for Multi‐Field Hydro‐Mechanical Consolidation Analysis

ABSTRACT Ground deformation induced by pore‐pressure dissipation poses significant challenges for geotechnical stability, yet traditional numerical solvers for Biot's consolidation often suffer from high computational cost and poor convergence, while standard PINNs struggle with gradient competition, inadequate spatial correlation modeling, and insufficient high‐order derivative accuracy. To address these limitations, this study proposes AGM‐PINNs, a novel framework that integrates a graph‐attention‐based spatiotemporal alignment module, an adaptive wavelet activation function (AWAF), and a mixture‐of‐experts (MoE) optimizer to efficiently solve three‐dimensional Biot's consolidation problems. The graph attention mechanism dynamically focuses on high‐gradient regions such as seepage fronts and stress concentration zones; AWAF enhances numerical smoothness and derivative fidelity; and MoE adaptively balances the multiphysics residuals to mitigate gradient competition. Extensive numerical experiments, including 1D, 2D, and 3D benchmarks, demonstrate that AGM‐PINNs achieve training losses on the order of , reduce displacement prediction errors from 8.954 mm in standard PINNs to 0.205 mm, and accurately infer permeability parameters with an average absolute error of . These results highlight the framework's robustness, high accuracy, and strong applicability to complex, multiscale hydro‐mechanical coupling problems, offering a reliable and mesh‐free computational tool for practical geotechnical engineering.