Machine learning for finite element analysis of fillet radius under tensile load

• Main Author: Goh, Nigel Jian Wei
• Other Authors: Chow Wai Tuck
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2020
• Subjects: Engineering::Aeronautical engineering
• Online Access: https://hdl.handle.net/10356/140541
• Institution: Nanyang Technological University

The improvement in both computational hardware power and software capabilities has enabled machine learning to take a more prominent role in today’s society. Across a myriad of industries like finance, engineering and the medical field, machine learning has increasingly become an effective tool. In particular, supervised learning allows for reliable predictions based on patterns gleaned from existing data. To do so, there are many algorithms that have been incorporated for supervised learning. This paper illustrates a unique approach of applying supervised learning with a shallow neural network to simplify the finite element analysis of a fillet radius under tensile load. The training is carried out using a relatively small set of 95 data samples using 2D linear elements with variations in the size of the radiuses and the inner areas of the mesh. These 95 samples are made up of coarse meshes with only 2 elements at their radiuses. The three backpropagation algorithms employed in this problem are the Levenberg-Marquardt algorithm, the Bayesian Regularization algorithm and the Scaled Conjugate Gradient algorithm, and the two transfer functions (or activation functions) evaluated are the pure linear transfer function and the tangent sigmoid transfer function. The target solution for the model is the highly refined finite element method solution for maximum stress value. Following this, the trained neural networks are applied to other alternative 2D and even 3D models. For the analysis of the other models, it was shown that the predictive errors are relatively low, with errors generally below 5%. The Levenberg-Marquardt algorithm was found to have the lowest errors for all the models. In contrast, FEM analysis for similarly coarse models yielded much larger errors of over 20% and refined models with 8 or 12 elements at the radius were required to obtain similar results to supervised learning. This result indicates that supervised learning has predictive capabilities that can potentially aid in finite element analysis for stress concentrations.