Information technology for predicting the hysteresis behavior of shape memory alloys based on a stacking ensemble machine learning model

Abstract

hape Memory Alloys are characterized by a nonlinear hysteretic behavior on the stress–strain (σ–ε) diagram, where the loop area determines the amount of energy dissipated per cycle. In this work, an ensemble Stacking machine learning model was developed to predict the hysteresis behavior of SMAs under cyclic loading at different frequencies (0.5, 1, 3, and 5 Hz). The model was constructed using experimental data from 100–250 loading cycles. Random Forest, Gradient Boosting, Extra Trees, k-Nearest Neighbors (kNN), Support Vector Regression (SVR), and Multilayer Perceptron (MLP) were employed as base algorithms. The ElasticNet model was selected as the meta-learner and tuned using GridSearchCV with GroupKFold validation. This approach ensured the combination of ensemble stability with adaptive selection of the most informative predictions from the base models. The obtained results showed a high accuracy in reproducing the stress–strain relationship: R² > 0.995, MSE < 0.0007, MAE < 0.02, and MAPE < 1.3 % on the test data. Validation on independent cycles 251 and 300 confirmed the model’s generalization ability, achieving R² > 0.974, MSE < 0.007, MAE < 0.06, and MAPE < 4.8 %. The interpretability of the model was provided by the SHAP method, which quantitatively determines the contribution of each input feature to the prediction. It was found that Stress is the dominant factor influencing the prediction, while UpDown defines the loading–unloading phase, and Cycle reflects the accumulation of cyclic effects. The developed ensemble Stacking model is an integral component of an information technology framework for predicting the hysteresis behavior of shape memory alloys using machine learning methods. The proposed approach provides not only high prediction accuracy but also a physically grounded interpretability of the results.