Active Learning: Accelerating Discovery of Optimal Optical Materials through Synergistic Computational Approaches

Inorganic crystals combining both a high refractive index and a wide band gap are of great importance for many advanced optical applications. However, the observed inverse relationship between these two properties hampers a straightforward identification of promising materials. To go beyond a costly and time-consuming trial and error approach, an ab initio high-throughput screening was performed on more than 4000 semiconductors [1]. Yet, this selection approach is limited to the compounds that are considered initially. It may therefore inadvertently overlook potentially valuable areas within the materials space.

To go one step further in this search, we here introduce an active learning framework powered by the synergistic use of an AI agent, open databases (via OPTIMADE [2]) and density-functional theory calculations. This framework strategically guides the exploration of crystal structures, emphasizing diversity within the chemical space and the exploitation of essential criteria for optical materials. Our findings demonstrate a substantial acceleration in comparison to conventional methods. The framework operates autonomously, eliminating the need for human intervention. Furthermore, we anticipate its potential to generalize to other material properties in the future.



[1] F. Naccarato et al., Phys. Rev. Mater. 3, 044602 (2019).

[2] Andersen, C. et al. Sci Data 8, 217 (2021).