物理信息神经网络在岩土工程中应用分析

打开文本图片集

中图分类号：TU43 文献标志码：A 文章编号：1672-1098（2025）05-0095-14

Abstract： Objective To verify the computational feasibility of Physics-Informed Neural Networks （PINN） under mesh-free conditions，clarify its applicability in complex geotechnical engineering mechanics problems，and provide theoretical and practical references for intelligent geotechnical computing methods，systematic application research of PINN in geotechnical engineering scenarios is caried out.Methods First， this study reviews the theoretical basis and algorithmic framework of PINN，and then conducts verification through three typical case studies： For the plane strain analysis of a double-layer earth-fill gravity dam，results show that PINN exhibits good consistency with the finite element method （FEM） in predicting displacement and stress distribution. Meanwhile， it demonstrates advantages in depicting boundary smoothness and stress concentration

Zones.For the prediction of soil mechanical responses under surcharge loading，PINN can effectively capture the laws of vertical displacement attenuationand stress concentration，showing sensitivity to boundary effects and local peaks.For the one-dimensional consolidation problem， PINN successully reproduces the evolutionary characteristics of pore water pressure dissipation over time.Results Comprehensive comparison results indicate that in geotechnical engineering numerical simulations，PINN not only can approximate the results of traditional methods but also has certain advantages in handling complex boundary conditions，heterogeneous materials，and parameter inversion problems.The analysis showed that the ReLU activation function is more conducive to describing the mechanical behaviors of heterogeneous materials compared with tanh，and insuficient network capacity willead to an accuracy botteneck. Conclusion The research content provides preliminary verification and methodological support for the popularization and application of PINN in geotechnical engineering.

Key Words： Physics-Informed Neural Networks; geotechnical engineering; machine learning

近年来，随着人工智能和深度学习技术的迅猛发展，物理信息神经网络（Physics-InformedNeuralNetworks，PINN）作为一种新兴的计算范式，为解决偏微分方程提供了一种全新的思路[1]。（剩余19086字）