Modelling of functional properties of SMA using machine learning methods

Abstract

This study deals with the modelling of NiTi shape memory alloy dissipated energy by means of supervised machine learning methods, considering the loading frequency. Shape memory alloys are materials of high interest both to science and industry. These materials are enjoying wide popularity due to their two peculiar properties: unique effect of shape memory and superplasticity, caused by direct austenite-martensite phase transformation and reverse martensite-austenite transformation. The traditional deterministic methods of assessment of material properties are often costly, time-consuming, and demand a well-trained workforce and laboratory equipment. On the contrary, in recent years, the methods of artificial intelligence have gained widespread attention due to their ability to reveal hidden insights from existing data. Machine learning is a subset of artificial intelligence. It allows training based on the available data and becomes better with time without the explicit need to be programmed. The experimental dataset was taken from open scientific sources. It contained the hysteresis curves for six loading frequencies of 0.1, 0.5, 1, 5, 7, and 10 Hz. The input data consisted of the next features: stress s (MPa), cycle number N, and loading frequency f (Hz). Based on these data, for each loading cycle, and for each loading cycle, the dissipated energy was calculated. To remove noise, Locally Weighted Scatterplot Smoothing (LOWESS) smoother in the nonparametric package of statsmodels was utilized. After that, the trapezoid numerical integration method was employed to calculate the area enclosed by the hysteresis loop of the respective cycle, that is, the dissipated energy. To augment the dataset, its points were interpolated using the modified Akima interpolation method (makima). Four models were built using the methods of Random Forest, AdaBoost, Gradient Boosting, and Neural Network. The best results were shown by the ensemble methods, such as AdaBoost, and Random Forest. For instance, the MAPE of AdaBoost was just 0.074, whereas the MAPE of Random Forest was 0.144. It was found that the Gradient Boosting method and Neural Network are not suitable for such a dataset, since the errors are quite large and, therefore, these methods are not good enough to be employed for solving such a problem.