Machine learning approaches to materials discovery have great potential, but currently face some limitations in data availability, curation, and potential biases. One promising approach is using generative machine learning models to produce new data points representing novel material compositions and structures. This can help address issues around data availability and fidelity.
In our work, we demonstrate an implementation of this idea to predict novel stable materials. Specifically, we train a generative model and use it to suggest the composition LiZn2Pt. We successfully synthesize this material, validating the prediction. This also allows us to extrapolate to other unreported ternary compounds in the same materials family.
Past machine learning work in this area has been constrained to known phase spaces and potentially reflected researcher biases. Our approach demonstrates that generative models can expand materials exploration into new spaces in an unbiased way. This could enable inverse design, where models accurately suggest materials compositions and structures that have desired target properties. This approach has significant implications for guiding experimental work towards the design of materials with specified real-world applications.