Machine learning metamodels for thermo-mechanical analysis of friction stir welding

This study explores the development and application of machine learning (ML) metamodels for the thermo-mechanical analysis of Friction Stir Welding (FSW). The main objective is to address the challenge of accurately predicting the thermo-mechanical behaviour of materials in FSW processes. Using finite element models, a high-fidelity dataset consisting of 20 Hammersley design datapoints is generated which is then used to develop a low-fidelity dataset of 420 datapoints using KNN (K-Nearest Neighbor) imputation. This low-fidelity dataset is used to train and test nine different ML metamodels (namely Linear Regression, Random Forest (RF), Support Vector Machines (SVM), AdaBoost, Gaussian Process, Gradient Boosting, Decision Tree, Histogram-based Gradient Boosting and Extreme Gradient Boosting). The performance of these metamodels is evaluated based on various metrics like \({R}^{2}\) $R^2$ (Coefficient of Determination), MAE (Mean Absolute Error) and MSE (Mean Squared Error). The findings reveal significant variance in the metamodels’ performance. Notably, Decision Tree, Gradient Boosting, XGB (Extreme Gradient Boosting) and Random Forest metamodels are found to be the top four performers.