Active learning-assisted search for thermal storage used TiNi shape memory alloys

TiNi-based shape memory alloys are promising candidates for thermal storage applications. However, a key indicator of thermal storage property, latent heat, is still less than desirable. Here, we use an active learning method with experimental feedback to guide the discovery of TiNi-based alloys with improved latent heat. The key features that affect latent heat are first screened out from a large feature pool, with which machine learning models are trained and applied to unknown alloys for predictions. We then use Bayesian optimization that considers both predictions and associated uncertainty to recommend alloys for experiments, and the results augment the initial data for next iteration. After four iterations, we successfully synthesized 15 alloys and one, Ti₂₅Ni_49.5Fe_0.5Hf₂₅, exhibits well-balanced latent heat and thermal hysteresis that outperforms reported ones. The designed alloys may find suitable thermal storage applications at elevated temperatures.