Machine learning predicted magnetic entropy change using chemical descriptors across a large compositional landscape

Abstract

Magnetocaloric refrigeration has drawn considerable attention in the last few decades as it can positively disrupt the current cooling technology. Most research efforts focus on developing new magnetic materials in the laboratory by trial and error. Here we report a materials dataset developed using past experimental work comprising several important magnetocaloric material classes such as La(Fe,Si/Al)13, heusler alloys, manganites, Gd5(Si,Ge)4 family, rare-earth and metallic glasses as well as Laves phase compounds with their reported magnetic entropy changes, -ΔSM(T,H). Notable linear and non-linear machine learning models are implemented to predict the -ΔSM(T,H) of materials. Our analyses indicate that the Random Forest model outperforms the others with R2 of 0.82. We then use this model to screen a large magnetic materials database with nearly 40,000 compounds to identify potential new magnetocaloric materials operating near room temperature. MnGa2Sb2, CrGa2Sb2, SbSCl0.1I0.9, Sm3Te4, LaRhSn, SbSI, Tl0.58Rb0.42Fe1.72Se2, Cs0.86Fe1.66Se2, La2.1MnGe2.2 are some of the newly predicted compounds that could yield large magnetocaloric cooling performance.