End-to-end prediction and design of additively manufacturable alloys using a generative AlloyGPT model

Abstract

Being able to tailor the composition at the voxel-size resolution, additive manufacturing of alloys calls for effective models to explore the vast and complex design space. We present AlloyGPT, a generative alloy-specific language model that concurrently performs forward property prediction and inverse alloy design. By converting physics-informed alloy data into structured textual representations, our model learns to capture intricate composition–structure–property relationships. It demonstrates high predictive accuracy across multiple phases and properties (R2 = 0.86-0.99) and robust generalization to unseen compositions. In inverse design tasks, it can generate diverse alloy candidates that meet specified property targets, showcasing its versatility. Comprehensive attention patterns and reasoning paths are observed within the model, suggesting promising clues for underlying alloy physics. By synergizing accuracy, diversity and robustness in prediction and design tasks, AlloyGPT is expected to accelerate knowledge integration and material design for uniform or gradient structural alloys manufactured by traditional and additive manufacturing.