Application of machine learning in efficient stress recovery in finite element analysis

Abstract

In this work, artificial neural network (ANN), a machine learning technique is applied in the field of stress recovery in finite element analysis (FEA). The most crucial design factor in structural analysis is stress. However, stresses calculated directly from FEA, are not accurate except at a few points called the super-convergent points. The stresses at the super-convergent points are targeted to map stress field using ANN for the whole domain. A higher-order element is used to calculate the reference stress values. The errors are calculated in the absolute sense and the results are compared with the superconvergent patch recovery (SPR) technique and also with the directly calculated stresses. Multiple mesh sizes are used to check the convergence with mesh refinement. The Cook’s skew beam problem is solved to apply the stress recovery using ANN.