An Explicit FEM-NN Framework and the Analysis of Error Caused by NN-predicted Stress

Shaoheng Guan,Y. T. Feng,Gang Ma,Tongming Qu,Mengqi Wang,Wei Zhou

DOI: https://doi.org/10.1007/s11440-023-01980-8

2024-01-01

Acta Geotechnica

Abstract:In the implicit FEM solver for nonlinear static problems, the tangential material matrices at each Gauss point are required to be accurately evaluated in the Newton–Raphson iterations. When phenomenological constitutive models are replaced by neural network-based models, the non-linear iteration may suffer from non-convergence when the neural network (NN) predicts the tangential matrix with insufficient accuracy. In this work, an explicit FEM solver coupled with a neural network-based constitutive model is proposed to bypass the need for the tangential matrix. The network is trained on the data generated from two constitutive models (IME model and CSUH model) respectively. This work presents macroscopic results and Gauss point stress–strain curves extracted from biaxial, retaining wall and rigid strip footing simulations. The trained models embedded in the FEM calculations are nearly able to reproduce the results of both classical constitutive model calculations at the macroscopic level, however, with visually higher errors at the level of Gauss points. Subsequently, a novel check-and-revision method is proposed to optimise the NN-based FEM computation iteratively by enriching the network’s training dataset and improving its generalisation on the Gauss point stress prediction. Compared with implicit FEM, explicit FEM eliminates the reliance on the tangential matrix, but the increasing loading steps and error accumulation problems pose a significant challenge to the prediction accuracy and robustness of the neural network. Inspection of the appearance and propagation of the prediction error yields that the best predictions at the Gauss point can not guarantee the NN-based computation's success, but the worst predictions lead to computational failure.