Inverse design of high-entropy superalloys using machine learning and generative artificial Intelligence

• End-to-end inverse design yields decision-grade HESA shortlists from data. • Creep (LMP) and oxidation ( log K A ) are optimized as first-class objectives. • Composition-only, physics-informed surrogates with descriptor and outlier control. • Feasibility floor + Pareto + Ward-medoids deliver a diverse experimental shortlist. • Conditional VAE explores new alloys under feasibility and sustainability filters. We introduce a structure-agnostic inverse-design workflow that turns heterogeneous literature and database evidence into experiment-prioritized shortlists of high-temperature high-entropy superalloy candidate chemistries. Unlike most data-driven high-entropy alloy (HEA) design studies that optimize a small set of proxies, we provide an end-to-end, reproducible decision-support loop that treats high-temperature creep and oxidation as first-class objectives and outputs compact shortlists for downstream validation. From curated multi-source data, we learn structure-agnostic, physics-informed surrogate models for the key high-temperature objectives – creep resistance (Larson–Miller parameter), oxidation kinetics (parabolic rate constant), melting point, density and elastic properties – and map predictions to “desirability” normalized scores. Candidate alloys are screened by a uniform feasibility floor and Pareto non-domination, then compressed by Ward–medoid clustering to yield compact, diversity-preserving shortlists for downstream validation. To explore beyond brute-force enumeration under the same admissibility rules, we couple the screening stage to a constraint-conditioned variational autoencoder, and retain only generated candidates that pass the full surrogate stack. The resulting compressed Pareto sets extend the occupied property envelope of legacy Ni-based superalloys while remaining interpretable through elemental-role analyses and Ashby-style trade-off maps. An external thermodynamic plausibility cross-check further shows that higher microstructure-oriented scores enrich the candidate space in single-phase HEA-compatible chemistries with wider stability windows and more favorable transition classes. Finally, we show how the same pipeline can be restricted to sustainability-compatible element pools, enabling performance-aware exploration under supply-risk and footprint constraints.