Interpretable machine-learning model for predicting the magnetocaloric effect in Heusler alloys

Heusler alloys are promising candidates to be used in magnetic refrigeration. This technology is friendly with the environment and relies on the magnetocaloric effect (MCE), the temperature change experienced by a magnetic material when a magnetic field is applied. In this work, a machine learning predictive model for the magnetic entropy change ( Δ S ) of Heusler alloys is developed. The dataset includes both, direct and inverse MCE Heusler alloys. The best performance of the predictive model is reached with the XgBoost model with determination coefficient of R v a l 2 ≈ 0 . 93 and R a l l 2 ≈ 0 . 97 . Explainable artificial intelligence (XAI) techniques were used to interpret the trained models and identify the descriptors that most strongly influence the predicted magnetocaloric response. The Ni–Mn–In, Ni–Mn–Ga, Ni–Mn–Sn families and some quaternary Heusler alloys are screened, identifying compositions with enhanced Δ S .