Discovering new electronic materials despite the small data problem

The search for new materials often starts with large data sets obtained from similar materials. But when searching for electronic materials that show abrupt changes in electrical resistivity, such as metal-insulator transition materials, there is not enough data to use standard machine learning techniques of discovery.

To address this small data problem, Wang et al. combined latent variable Gaussian processes, multiobjective Bayesian optimization, and high-fidelity electronic structure calculations. The resultant adaptive optimization engine accelerated the design of metal-insulator transition materials. They discovered 12 functional and synthesizable compounds in the lacunar spinel family on the design Pareto front within tens of iterations.

“Our work disrupts the machine learning status quo in materials science by building predictive and explorative models starting from a small data set without requiring either large data sets or hand-tuned descriptors,” said author James Rondinelli.

Eleven of the 12 discovered compounds have not been previously reported. The ground states of the 12 compounds are semiconducting, but they also exhibit two electronic transitions: the metal-to-insulator transition, which the authors expected, and the semiconductor-to-insulator transition, which the authors did not expect. Two of the compounds, InWMo₃Se₈ and InTaMo₃Se₈, exhibited promising simulated DC electrical resistivities for functionality.

The authors will work with their collaborators to synthesize and test the 12 compounds they discovered as well as extend their method beyond the lacunar spinel family. They believe their method could benefit material searches when generating data is time or cost intensive.

Source: “Featureless adaptive optimization accelerates functional electronic materials design,” by Yiqun Wang, Akshay Iyer, Wei Chen, and James M. Rondinelli, Applied Physics Reviews (2020). The article can be accessed at https://doi.org/10.1063/5.0018811 .