Structure-to-process modeling drives experimentally validated unified dual-phase steel

Abstract

Unified dual-phase (UniDP) steels enable tailored performance from a single composition, revolutionizing sustainable material systems by addressing recyclability and weldability challenges. Traditional design frameworks, constrained by forward "process-structure" models and costly uncertainty quantification, falter under sparse data and complex microstructures. Here, a microstructure-centric inverse design strategy is proposed that replaces uncertainty quantification with direct "structure-to-composition/process modeling", leveraging real microstructural features to map composition and processing parameters. Specifically, our approach integrates a variational autoencoder to encode authentic microstructural features into a latent space and a multilayer perceptron to predict composition, processing routes, and properties. Combined with specific latent space sampling, the framework achieves high-efficacy design exploration. The experimental success of UniDP steels stands as a cornerstone of this work: the designed alloy consistently achieves the target properties in all three performance tiers, at a lower cost than other commercial alloys. Latent space analysis further validated the model's ability to interpolate seamlessly between microstructures and encode multi-scale property relationships, confirming its robustness for real-world applications. By experimentally demonstrating the viability of microstructure-driven inverse design, this work not only resolves longstanding barriers in complex alloy systems but also establishes a replicable, uncertainty quantification-free framework for sustainable material innovation.