Abstract
Finite element analysis (FEA) is used to solve real life engineering problems because it can model complex geometry and capture local stress concentration effects. Given the developments in machine learning due to advancements in both computational hardware technology and software capabilities, it has produced reliable predictions based on patterns from existing data. This paper intends to analyse whether machine learning can increase the efficiency for the finite element analysis of rivet holes under biaxial load. This is done by training the neural network with the coarse mesh of a rivet hole under biaxial load against its analytical solution. The training is carried out using a small set of 125 data samples using 3D linear elements with variations in inner radius, outer radius, and the inner thickness. The neural networked will be trained based on three common backpropagation network algorithms. They are the Conjugate Gradient Descent, Levenberg-Marquardt algorithm, and Bayesian Regularization. In addition, tangent sigmoid and pure linear transfer functions will be evaluated. The trained neural networks are then applied to the rivet hole model and the target solution for the model is the maximum stress value for the fine mesh model. It is shown that the predictive errors are relatively low with errors generally below 5%. The Levenberg-Marquardt algorithm using Tangent Sigmoid transfer function was found to have the lowest errors. This indicates that supervised learning has predictive capabilities to aid in finite element analysis for stress concentrations.
Keywords: Finite Element Analysis ∙ Computational Structural Analysis ∙ Supervised Learning
