Prediction of narrow HT-SMA thermal hysteresis behaviour using explainable machine learning

The ability to tailor narrow high-temperature thermal hysteresis behaviour in Ti-Ni-based high-transformation temperature shape memory alloys (HT-SMAs) is crucial for the development of the next-generation actuators intended for use in extreme conditions. Equally, the need and the development of tools to accurately predict shape memory alloys (SMA) transformation temperatures also becomes emergent. As such, the purpose of this study is to curate a small dataset from the experimental peer-reviewed literature and develop a practical data-driven machine learning (ML) based model to predict narrow hysteresis in HT-SMA transformations. This study compiles a small dataset based on experimental data and develops at least twelve candidate models using common machine learning and artificial neural network algorithms. Our results show that the XGBoost-based model exhibits the best performance in predicting the narrow thermal hysteresis behaviour in Ti-Ni-based HT-SMAs (i.e., the highest R2 = 0.893 and the lowest RMSE = 5.4). Further, using preselected case studies, the study demonstrates the utility of the relative feature importance, the accumulated local effects (ALE), and the SHapley Additive exPlanations (SHAP) model-agnostic techniques to interpret predicted outcomes and establish a more transparent relationship between the input materials composition descriptors and the predictions. These analyses demonstrate that the explainable ML approach offers practitioners a more transparent design tool, not only to predict the narrow hysteresis behaviour in HT-SMA transformations, but also a deeper understanding of how the input materials composition descriptors can contribute to its magnitude and/or directionality of the narrowing hysteresis.