Convolutional Neural Network Algorithm for Stress Field Approximation for a Uniaxially Loaded Plate with Holes

Abstract

ABSTRACT Knowing the high stress concentration and its location around a hole is of practical importance in designing engineering structures, effective parameters can be selected in order to achieve minimum stress concentration around the hole(s). The visualisation of the stress distribution is done experimentally or numerically using finite element analysis (FEA). This study involved the approximation of the von Mises stress field distribution around regular holes in finite metallic plates, based on deep learning 2D encoder-decoder convolutional neural network (CNN). The CNN based stress field prediction is achieved by nonlinear 2D image training data obtained from ABAQUS, the geometric parameter of flat plate and the hole(s) are the input to CNN and the von Mises stress filed distribution is the output, assuming the plate considered is finite, isotropic, plane stress state, linearly elastic and uniaxial loading condition with total number of trainable parameters for the neural network is 3,437,143. The prediction accuracy achieved is 94.18% with a loss of 0.03% between the ground truth and the predicted images Considerable care was taken to minimize the complexity of the CNN architecture and to make it interpretable and to achieve high accuracy with available low data. The proposed method can predict acceptable accurate solution to a problem with the geometries not included in the training dataset. This approach can be directly applied as it provides immediate feedback for real-time design iterations at the early stage of design. Keywords: Convolutional neural network, finite element analysis, isotropic plate, machine learning, regular hole, stress concentration, uniaxial loading