Traditionally, designing novel materials involves exploring new compositions guided by insights from previous work, relying on a trial-and-error approach, where continuous synthesis and characterization proceed until the properties meet the improvements. This method is inefficient due to the challenges of exploring vast chemical spaces. In this study, a machine-learning-based methodology is developed to assist the design from available data in the literature, allowing us to test in silico more than 1.2 million compositions. Two databases with 1227 inputs were created from published studies. Four machine learning (ML) models were trained over the feature sets using 517 compositional features (generated from 58 atomic properties) to predict magnetocaloric properties of perovskites: Curie temperature (TC), magnetic entropy change (ME), and relative cooling power (RCP). The best model-feature combinations were used to explore the chemical space of lanthanum, praseodymium, and neodymium manganites, identifying composition trends for different temperature applications, including room temperature refrigeration, where the most suitable combinations of doping elements were highlighted. The study offers valuable guidelines for future research insights on magnetocaloric materials, and the methodology can be transferred to other perovskite related material areas, such as catalysts and solar cell materials.
Read full abstract