An optimised stacking ensemble machine learning model for predicting the hysteresis behaviour of shape memory alloys

The study investigates the possibility of predicting the hysteresis behaviour of a NiTi shape memory alloy (SMA) under cyclic loading using machine learning methods. Based on experimental test data obtained from SMA wire specimens, an ensemble model of the Stacking Regressor type is developed, combining the Random Forest, Gradient Boosting, Extremely Randomised Trees, k-Nearest Neighbours, Support Vector Regression, and Multi-Layer Perceptron algorithms. The ElasticNet algorithm is employed as the meta-model. The ensemble structure and model hyperparameters are optimised using the GridSearchCV procedure with group cross-validation. The results demonstrate high accuracy in strain prediction across different cyclic loading frequencies. The coefficient of determination R² for the test data exceeds 0.996. The proposed model also exhibits strong generalisation when predicting material behaviour beyond the training range and accurately reproduces the shape of hysteresis loops in the stress-strain diagram.