Machine Learning Maps Predict the Possibility of Over 3,000 New Material Stages

In collaboration with the University of Tokyo (Yutokyo), the National Institute of Advanced Industrial Science and Technology (AIST), Tohoku University, Kyoto Institute of Technology, and the National Institute of Materials Science (NIMS), we have developed and published an “elemental reactivity map” to discover new phases.

The researchers proposed a map that uses machine learning to identify more than 3,000 element combinations. This could form a new phase from one of a total of 85,320 combinations of up to three elements selected from 80 easily processed labs. The results of this study were published in Materials Chemistry.

Inorganic materials are synthesized by reacting multiple elements. If a new material is successfully synthesized by an unprecedented combination of elements, and the phase exhibits special physical properties or useful functions, it can become a “treasure” that can actually be used as a new material.

However, many of the combinations present in the crystal structure database are combinations that have been previously attempted but simply failed to react, and the ability to predict synthesisability beforehand is a key factor in the efficient discovery of new phases.

The research team has developed 80 “elemental reactivity maps” in a 80 x 80 grid format. This indicates the possibility of phase formation from a combination of up to three different elements, along with the presence or absence of known materials.

These maps were created through machine learning using crystal structure data from over 30,000 inorganic compounds and are published as interactive web systems that are accessible to anyone.

When predictive results were verified using an experimental crystal structure database containing data for complex crystals and solid solutions, known compounds could be 17 times higher among combinations with high reactive scores (≥0.95) compared to combinations with low reactive scores (<0.05), indicating the validity of the reactive score.

With over 3,000 combinations of elements that exhibit high responsiveness scores but not present in the experimental database, the map is expected to serve as a “treasure” with hidden new phases.

Additionally, by actually utilizing these maps, the researchers have successfully discovered dozens of new stages, including the B20 structural alloy Co (AL, GE), which has attracted attention as a potential magnetic silmion or thermoelectric material.

By utilizing these elemental reactivity maps, various new phases are expected to be discovered, and new materials may be found between them. Furthermore, combinations of elements that are unlikely to react can also be identified from these elemental reactive maps, which can prove useful in identifying potential containers or electrodes that remain chemically inert.