Transformer-based multimodal learning for predicting mechanical properties in heat-treated stainless steel

Abstract

Accurately predicting mechanical properties of heat-treated materials is critical for intelligent process control and advanced manufacturing. This study proposes a Transformer-based multimodal learning framework for predicting the hardness and wear behavior of carburized steel after vacuum carburizing. By integrating microstructural images, material compositions, and process parameters, the proposed model effectively captures complex cross-modal relationships. Experimental results show that the multimodal model achieves high prediction accuracy, with an R² of 0.98 and MAE of 5.23 HV for hardness prediction. Furthermore, Variational Mode Decomposition (VMD) is introduced to preprocess the wear curve, reducing noise and improving the robustness of friction performance prediction. The results demonstrate the effectiveness and generalizability of the proposed approach, offering a practical AI-based solution for intelligent material property evaluation and process optimization.