Inverse Design of Materials That Exhibit the Magnetocaloric Effect by Text-Mining of the Scientific Literature and Generative Deep Learning

Magnetic materials play an important role in a wide variety of everyday applications, and they are critical components in many devices used for energy conversion. However, there are very few materials known to exhibit magnetism of any kind, and the slow process of experimentally driven magnetic-materials discovery has limited the development of devices for functional applications. In this work, a complete magnetic-materials discovery pipeline is presented that uses natural language processing (NLP), machine learning, and generative models to predict ferromagnetic compounds in the Heusler alloy family. Using the "chemistry-aware" NLP toolkit, ChemDataExtractor, a database of 2910 magnetocaloric compounds is autogenerated by sourcing from the scientific literature. These data are then used to train property-prediction models for key figures of merit that describe the magnetocaloric effect. The predictive models are applied to novel Heusler alloy material candidates that have been created using deep generative representation learning. Convex-hull meta-stability analysis and ab initio validation of these candidates identify six potential materials for solid-state refrigeration applications.