Abstract
We develop a transferable framework for microstructure- and chemistry-aware, explain- able prediction of tensile properties across metallic systems. Using peer-reviewed stain- less steel literature, we curate a dataset in which optical microscopy (OM) and scanning electron microscopy (SEM) images provide microstructural information that is comple- mented by compositional data to predict yield strength, ultimate tensile strength and elongation at fracture. Diverse image-based descriptors are used to extract information on microstructure and texture from micrographs, and model performance is evaluated both in-distribution via cross-validation and out-of-distribution via an independent ex- perimental dataset. Feature importance analyses identify the dominant compositional and microstructural factors controlling the investigated properties, and comparison with established metallurgical knowledge supports the physical plausibility of the learned relationships. Finally, we demonstrate the transferability of the proposed approach by applying instance-reweighted transfer learning from stainless steels to a dual-phase steel literature dataset, showing that the framework remains stable under domain shift and is suitable as a general template for composition–structure–property mapping from both OM and SEM micrographs.

This publication has 42 references indexed in Scilit: