Metallurgy-guided bidirectional generative framework for performance prediction and inverse process design in hot-rolled steel

To effectively handle the complex industrial data characteristics inherent in hot-rolled steel manufacturing, this paper proposes a Cycle-Consistent Bidirectional Generative Framework with Metallurgical Priors (C2-BIGF). Focusing on hot-rolled steel strips, the framework leverages an enhanced variational autoencoder (VAE) integrated with metallurgical priors for informed feature engineering and representation learning. Key process features are quantitatively extracted via physically-based microstructural evolution equations and are systematically integrated into both model input and solution validation, thereby embedding metallurgical prior knowledge throughout the entire modeling framework. The proposed approach adopts a bidirectional generative structure combined with a cycle-consistency mechanism, significantly enhancing prediction stability and structural coherence between forward property prediction and inverse process generation. Comprehensive validation is conducted using real-world industrial datasets, including rigorous ablation studies evaluating individual module contributions. Additionally, two practical industrial scenarios are presented: inverse generation of process parameters from desired mechanical properties under fixed alloy compositions, and cost-optimized alloy composition design under fixed process constraints. In each scenario, the feasibility and rationality of generated solutions are critically evaluated through metallurgical prior knowledge and cycle-consistency verification. Experimental results validate the framework's effectiveness. In forward prediction, it achieves an overall R² of 0.9689. In inverse design, it reliably generates solutions for target yield strengths within a ± 15 MPa margin and successfully produces cost-optimized alloy designs, effectively supporting customized and flexible steel production.