Lattice-dependent formation enthalpy enhances machine-learning phase prediction of high-entropy alloys

Abstract : Machine-learning (ML) models for high-entropy alloys (HEAs) phase classification commonly rely on empirical descriptors such as the mixing enthalpy ( ), which lack the dependence of lattice structure. Driven by the hypothesis that phase competition is fundamentally governed by lattice-specific energetics, this study proposes a conceptual framework replacing the lattice-agnostic with a Lattice-dependent descriptor—the difference between solid solution formation enthalpy of FCC and BCC structure ( ). We evaluate its effectiveness against conventional features across four classification schemes. Comparative ML analyses across four classification schemes show that provides accuracy comparable to conventional feature sets, while permutation-importance and SHAP analyses reveal that becomes the dominant predictor whenever FCC/BCC competition is involved, forming a coupled decision axis with VEC and revealing quantitative tipping points that delineate the FCC–BCC phase boundary. The SHAP–GAM analysis and external experimental data further reveals quantitative tipping points: FCC stability is favored at VEC > 8.37 and < –7.63 kJ/mol, whereas BCC stability emerges at VEC < 7.89 and > –3.11 kJ/mol. Guided by these criteria, two HEAs were designed and experimentally confirmed to form single BCC (Al 20 V 20 Mn 20 Fe 20 Co 20 ) and FCC (V 10 Mn 25 Co 25 Ni 40 ) phases, validating the mechanistic interpretation. These results establish lattice-dependent formation enthalpy as a meaningful thermodynamic parameter for understanding solid-solution formation in HEAs and provide a physically interpretable route for structure-guided alloy design.