Multi-fidelity, active-learning assisted design of TaNbMoVW refractory high-entropy alloys with superior mechanical properties

Abstract Refractory high-entropy alloys (RHEAs) are promising for extreme environments, yet the vast non-equiatomic design space hinders discovery of compositions with tailored strength-ductility balance. Here we develop a neuroevolution potential (NEP)-centered, multi-fidelity design framework for Ta–Nb–Mo–V–W RHEAs. The NEP achieves DFT-level accuracy for energies and forces, and its rapidly fitted Young’s modulus agrees with experimental benchmarks. SVR surrogates trained on NEP-labeled elasticity data enable rapid screening of more than 50 thousands compositions within seconds, NEP provides medium-fidelity validation within minutes, and tensile MD offers high-fidelity verification from 300 K–1200 K. Guided by different performance demands, the workflow identifies Ta 0.20 Nb 0.05 Mo 0.35 V 0.05 W 0.35 with superior strength and Ta 0.20 Nb 0.06 Mo 0.34 V 0.34 W 0.06 that sustains a combined strength-toughness advantage over equiatomic TaNbMoVW across 300 K–1200 K. The approach accelerates discovery by coupling models and simulations with complementary speed and fidelity, enabling scalable, data-efficient screening, reliable validation, and temperature-aware optimization across complex compositional spaces.