Mat-GPT Estimation of Material Properties

Accurate, early prediction of metal mechanical, electrical, and thermal properties is essential for material selection and process design, yet conventional characterization employed in traditional metal manufacturing process is costly and slow. Even after the recent development in Machine Learning (ML), there isn't a fully ML based software solution that exists for predicting such properties. We propose a transformer-based, multi-task model that is capable of predicting properties like Young's modulus and ultimate tensile strength directly from routinely available manufacturing inputs-chemical composition, process parameters, and the ordered sequence of treatments. By representing these heterogeneous, interdependent values as a table of numeric and textual tokens, multi-head attention learns intrinsic interrelations of these input components without hand-crafted features. The approach enables fast, non-destructive property estimation and reduces experimental burden. We outline data representation, model architecture, and evaluation protocols. This work highlights a practical path to software-based prediction of metal properties using only inputs available in production logs.