The TRIP-active composition domain in CoCrFeNiCu alloys: machine learning screening and atomistic mechanisms

To improve the strength–ductility synergy of CoCrFeNiCu alloys, a composition design strategy coupling molecular dynamics (MD) with machine learning (ML) is proposed. A dual-objective XGBoost model trained on 600 MD tensile cases (298–873 K) reached cross-validated R 2 of 0.957 for strength and 0.821 for ductility, and three rounds of active learning (prediction deviation ≤ 0.41%) identified three optimized compositions on the Pareto front. The two high-ductility alloys share a common signature: Co enrichment to 31.9–35.2 at.% and Cu depletion to 5.7–12.0 at.%. Relative to the equiatomic baseline they show 27–28% higher ductility and a 32% higher strength–ductility product over 298–873 K. The gain arises from a deformation-induced FCC-to-BCC shear transformation that begins within the uniform deformation stage, at onset strains of 0.139–0.140 against 0.172–0.179 for the baseline, and reaches peak BCC fractions of 38–54% against 9–13%. The transformation restores work hardening after the initial stress drop, which the baseline alloys do not exhibit. Nucleation proceeds by stress-induced homogeneous shear without a dislocation precursor, and the transformation reverts almost completely on unloading via Shockley partials. The bimodal ductility distribution predicted by ML corresponds strictly to this mechanistic dichotomy, showing that the model locates metastable composition domains without prior mechanistic input.