Knowledge-Graph-Guided Design of Compositionally Flexible Alloys for Sustainable Metal Recycling

Developing engineering alloys that retain target properties despite compositional variability in mixed metallic scrap remains a major challenge in sustainable alloy design. Machine learning can accelerate alloy development, but limited datasets, extrapolation into underexplored composition spaces, and insufficient physical constraints reduce prediction reliability. To address these challenges, a recycling-oriented framework based on physicalmetallurgy knowledge graphs, termed ReAlloy-KG, is developed. Four task-specific models predict martensite start temperature (Ms), stacking fault energy (SFE), yield strength (YS), and pitting potential (Epit), and their outputs are integrated for multi-objective screening. Physically constrained screening of 20,301 compositions generated by virtual blending of 321 and 201LN stainless steels with Fe-50Ni (wt.%) powder feedstock identifies a continuous admissible region that contains 1,995 candidates and satisfies phase stability, deformation stability, corrosion resistance, and yield strength criteria. Experiments on five boundary-selected alloys confirm a stable face-centered cubic austenitic matrix before and after tensile deformation, yield strengths near the target range, and pitting potentials above both the screening threshold and the commercial 304 stainless-steel benchmark. This work establishes a transferable strategy for designing compositionally flexible alloys capable of accommodating the unavoidable compositional variability inherent in mixed scrap streams.