Teaching Solid Mechanics to Artificial Intelligence: a fast solver for heterogeneous solids

TL;DR

This paper introduces a deep neural network surrogate model that predicts local stresses in heterogeneous solids with high accuracy and significantly faster than traditional solvers, applicable to both elastic and elasto-plastic materials.

Contribution

The paper presents a novel DNN-based surrogate model capable of rapid stress prediction in complex heterogeneous materials, outperforming standard iterative solvers in speed and accuracy.

Findings

Achieves about 3.8% MAPE in elastic media

Performs 103 times faster than spectral solvers

Maintains accuracy with 6.4% MAPE in non-linear elasto-plastic cases

Abstract

We propose a deep neural network (DNN) as a fast surrogate model for local stress (and in principle strain) calculation in inhomogeneous non-linear material systems. We show that the DNN predicts the local stresses with about 3.8% mean absolute percentage error (MAPE) for the case of heterogeneous elastic media and a mechanical phase contrast of up to factor of 1.5 among neighboring domains, while performing 103 times faster than spectral solvers. The speed-up arises from the fact that after training, the DNN predicts the stress without any iterations, as opposed to the iterative nature of standard non-linear solvers. The new DNN surrogate model also proves suited for general purposes: it is capable to reproduce the stress distribution in geometries topologically far different from those used for training, implying effective learning of scenarios described by the underlying partial differential equations. Even in the case of elasto-plastic materials with up to 2 times mechanical contrast in elastic stiffness and 4 times in yield stress among adjacent regions, where conventional solvers typically require a substantial number of iterations to arrive at stress predictions, the trained model simulates the micromechanics with a MAPE of 6.4% in one single forward evaluation step of the network, i.e. without any iterative calculations even for the case of such a non-linear problem. The results reveal a completely new and highly efficient approach to solve non-linear mechanical boundary value problems and/or augment existing solution methods in corresponding hybrid variants, with an acceleration up to factor of 8300 for heterogeneous elastic-plastic materials in comparison to the currently fastest available solvers.