A Multimodal Dynamic Cross‐Attention Feature Fusion Network for Mechanical Property Prediction of Hot‐Rolled Steels

Mechanical properties of steels depend on their internal microstructures. However, current mechanical property prediction methods mainly rely on composition and process data, making it difficult to fully reflect the effects of microstructure on mechanical properties. Thus, the MDCAFF‐Net (multimodal dynamic cross‐attention feature fusion network) was proposed. The MDCAFF‐Net employed a dynamic cross‐attention feature fusion module (DCAFFM) and a multiscale feature attention fusion module (MSFAFM) to improve the prediction accuracy of the model. Compared to the bagging regressor (BR), random forest (RF), and deep neural network (DNN), the MDCAFF‐Net showed lower mean absolute error (MAE), mean absolute percentage error (MAPE), and root mean squared error (RMSE) and higher R 2 values for yield strength (YS), tensile strength (TS), and elongation (EL). Furthermore, compared to the fully connected layer data concatenation method (FCLDCM), the MDCAFF‐Net model for YS, TS, and EL showed improvements of 13.27%, 11.9%, and 25.38% in R 2 , decreases of 59.42%, 49.21%, and 79.69% in MAE, lower MAPE by 60.58%, 48.53%, and 79.33%, and a reduction in RMSE by 55.24%, 46.51%, and 60.92%, respectively. As a result, the proposed MDCAFF‐Net achieves precise predictions of mechanical properties based on multimodal data.