Physics-Informed Machine Learning for Refractory Alloy Design

Abstract

Refractory complex concentrated alloys (CCAs) exhibit exceptional mechanical properties at elevated temperatures, but their vast compositional space (approximately 10^15 possible combinations) poses significant challenges for experimental exploration. We develop a physics-informed machine learning framework combining quantile gradient boosting with Born stability constraints to predict eight mechanical properties (bulk modulus B, shear modulus G, Vickers hardness Hv, elastic constants C11, C12, C44, Young's modulus E, and Poisson's ratio nu) of refractory CCAs. Using 393 density functional theory (DFT)-computed alloys from a 10-element design space (Cr, Hf, Mo, Nb, Re, Ta, Ti, V, W, Zr), we achieve coefficient of determination (R^2) values of 0.89 to 0.97 with mean absolute errors of 0.85 to 14.74 GPa across six properties on a held-out test set (n = 59). Remarkably, all test predictions satisfy the Born stability criteria (100% compliance), demonstrating the efficacy of physics-informed constraints. Compositional screening in valence electron concentration (VEC)-atomic size mismatch (delta) space identifies high-performance candidates, with MoReW exhibiting the highest predicted shear rigidity (C44 = 139 GPa). Elemental analysis reveals that rhenium appears in 100% of the top-performing alloys, molybdenum in 90%, and chromium in 80%, establishing quantitative design rules: VEC = 6.0 to 6.4, delta < 10%, and Re-Mo-Cr ternary systems maximize shear resistance. This physics-informed screening framework enables accelerated discovery of mechanically stable, high-performance refractory alloys from the trillion-scale compositional space.