Facilitating creep-resistant Al alloy design through a physics-aware integrated machine learning pipeline

Creep-induced failures threaten the reliability of Al-based hot-end components. Despite advances in ML-aided design of heat-resistant Al alloys that have alleviated experimental and modeling burdens, creep tolerance metrics are sporadically distributed across heterogeneous modalities, and conventional models still struggle with limited physical interpretability due to the single composition–property mappings. Here we develop a hybrid machine learning pipeline, incorporating LLM‑based data mining and high‑throughput CALPHAD, to predict the creep threshold stress ( σ th ) of cast Al alloys with insight-encoded representations. Within the proposed Al–Ni–Er–Zr–Sc–Mn system, a family of candidate compositions is identified with predicted threshold stresses greater than 70 MPa at 573 K under the dislocation‑climb mechanism. This 11% improvement over the upper limit in our curated database results from three synergistic steps: high-fidelity meta-feature extraction, infusion of domain-informed descriptors, and explicit feature dimensionality reduction. Moreover, explicable strategies highlight interphase misfit strain, volume fraction of L1 2 precipitates, diffusional conductivity, and volume fraction of Al 3 Ni microfibers as four dominant contributors to thermodynamic and kinetic stability effects on the threshold. Leveraging these factors in assembling objective functions, composition screening via high-throughput CALPHAD converged on optima matching those from Bayesian optimization on the ML surrogate. The dual authentication between physics-aware and data-driven searches pioneers a reproducible pathway to efficiently develop creep-resistant lightweight alloys for next-generation propulsion and energy systems, where pistons and cylinder heads in diesel or downsized engines and related hot-zone castings require long-term dimensional stability.