Action space and features for complex multicomponent alloys and ceramics property prediction with deep reinforcement learning

Multicomponent alloys and ceramics including alloys for thermoelectrics, ultra-high-temperature ceramics for aerospace applications, and refractory high entropy alloys (HEA) have seen rapid growth due to their superior mechanical, thermal, and magnetic properties. The large configurational space offers unique opportunities for discovery of new configurations with wide range of physical properties. Describing the unique atomic environments is crucial for establishing material-property relationship. We develop and implement a set of descriptors based on the elemental properties and ordering of atomic species. In combination with the actions derived from the engineered descriptors we propose a deep reinforcement learning approach to reach the target property. We demonstrate the approach to predict physical properties for discover configurations for ternary compounds such as (Al_xGa_yIn_z)_2NO_3N and three refractory HEAs: MoNbTaW, MoNbTaVW, and MoNbTaTiW.