Navigating materials design space with variational autoencoders to learn materials thermodynamics

Recent advances in artificial intelligence for materials design and discovery target the screening of entire libraries of materials for desirable properties as well as the prediction of novel materials for target properties. A particular challenge has been the materials design for thermodynamic properties far away from zero Kelvin and ambient pressure because of the lack of public thermodynamic data.

To overcome this data limitation, here we leverage chemical/materials similarities to predict the thermodynamic properties of similar materials. Our approach makes use of variational autoencoders (VAEs) and extensions/modifications thereof that learn a compact and continuous low-dimensional latent space from an input materials fingerprint. Such models have previously demonstrated capabilities for materials reconstruction, generation and interpolation. We introduce a definition of materials similarity in (property-ordered) low-dimensional latent spaces, compare it to the materials similarity in the input materials fingerprint space, and demonstrate transferable models of thermodynamic properties.