Property-aimed embedding: a machine learning framework for material discovery

Proposing new materials by atom substitution based on periodic table similarity is a conventional strategy of searching for materials with desired property. We introduce a machine learning frame work that promotes this paradigm to be property-specific and quantitative. It is of peculiar usefulness in situations where abundance data is accessible for learning general knowledge but samples for the problem of interest are relatively scarce. We showcase its usage and viability in the problem of separating high entropy alloys with different structural phases, for which a very simple data-driven criterion achieves differentiating ability comparable with widely used empirical criteria. Its flexibility and generability make it a promising tool in other material discovery tasks and far beyond.