Machine learning-powered modelling of creep strain in shape-memory Nitinol alloys at different combination points of stress and temperature

Abstract

An integrated machine learning system has been developed to make creep strain predictions on shape-memory Nitinol alloys subjected to different stress ranging from 0 to 500 MPa and temperature ranges from 25°C to 80°C. Our work used experimental data from 35 strain-controlled tests through which we evaluated four machine learning approaches, including linear regression and k-nearest neighbours along with decision Tree and Random Forest because traditional methods showed insufficient capabilities for analysing stress/temperature/creep deformation relationships. The evaluation of the models occurred through the mean absolute error(MAE) along with the root mean square error (RMSE) and the coefficient of determination (R²). The Random Forest Ensemble Method exceeded the performance of linear models and Decision Tree models by providing near-perfect predictive power with R² = 1.000 while maintaining a zero MAE = 0.000 and RMSE = 3.93 × 10⁹. This matched better than linear models (e.g. linear regression: R² = 0.708, MAE = 0.009) and decision tree (R² = 0.999, MAE = 0.000). Stress exists as the main cause of creep strain based on observation because there is a positive correlation (r > 0.95) while temperature maintains a non-linear effect that hastens deformation rates to 40% at 80°C when compared to 25°C. This study leads the field by merging machine learning algorithms with thermomechanical Nitinol data to build a comprehensive predictive system that increases material reliability and optimisation potential for biomedical and aerospace and energy sectors.