Closed-loop inverse design of high entropy alloys using symbolic regression-oriented optimization

Abstract

Rapidly finding new materials that are distinct to those in existing datasets continues to be a challenge for machine learning-driven approaches. Here, we propose a closed-loop framework to accelerate the inverse design of target materials, with emphasis on the use of symbolic regression-guided optimization. The refractory high entropy alloys are used as a model system to demonstrate the efficacy of the proposed approach. Symbolic regression learns a simple formula between a basic physical descriptor (enthalpy of fusion) and target property (yield strength at 1000 °C), which allows us to devise a new alloy system (V-Ti-Mo-Nb-Zr). The property optimization is enabled by combining heuristic algorithms and an uncertainty-aware utility function to recommend candidates for experiment. With only four iterations, we fabricate 21 alloys, of which 12 exhibit improved specific yield strength and two surpass 110 MPa/(g/cm${}^3$). The gradual rise in density coupled with the quick increase in lattice distortion underpin the enhanced yield strength. This study highlights the effectiveness of symbolic regression-oriented optimization in identifying target materials from complex systems.