Data-driven design of Nb-W refractory alloys using Transformer-based stress-strain modeling

Generative AI provides a powerful route for predicting material behavior in data-scarce regimes, an especially pressing challenge for refractory alloys where extreme processing environments and high melting points hinder large-scale data collection. Here, we introduce a Transformer-based generative framework that models the entire deformation trajectory rather than focusing solely on isolated scalar properties (such as YS, UTS), by synthesizing full stress–strain curves of Nb-W-based refractory alloys. Applied to Nb521-derived systems, the model captures the sequential evolution of tensile deformation and accurately predicts elastic–plastic responses across a wide compositional and processing space. The framework demonstrates utility in three downstream tasks: (i) inverse design of composition and processing, yielding >35% strength improvement without loss of ductility; (ii) interpretable regime analysis via a Mixture-of-Experts (MoE) mechanism, highlighting the beneficial roles of Hf and Ta in ductility, consistent with experimental and DFT validation; and (iii) seamless integration of generated curves into finite element simulations for structural performance assessment. Together, these results establish a coherent generative-analytic-predictive workflow for mechanism-aware, data-efficient design of advanced refractory alloys.