In this study, an optimized deep learning (DL) framework was developed to predict the mechanical properties of dual‐phase 980 MPa (DP980) steel from scanning electron microscope (SEM) microstructure images. A sample database was constructed through heat treatment experiments, and the DL model was optimized via hyperparameter tuning. The best performance was achieved with six convolutional layers, yielding root mean square errors (RMSE) of 6.98 MPa for ultimate tensile strength (UTS) and 0.53% for total elongation (TE), with R 2 values exceeding 0.90. Subsequently, the mechanical properties of DP980 steel under various heat treatment conditions were predicted using the DL model based on SEM images. The results showed that the prediction errors for UTS and TE were within 60 MPa and 1.2%, respectively. To enhance the interpretability of the model, the Gradient‐weighted Class Activation Mapping (Grad‐CAM) algorithm was employed to visualize the relationship between microstructure and mechanical properties. The results revealed a strong correlation between UTS and martensite due to its high strength, and between TE and ferrite due to its high deformability. When ferrite was constrained by surrounding martensite, TE also exhibited partial dependence on martensite phase. This work confirms the capability of image‐based deep learning for property prediction and supports microstructural design and optimization.
Ya et al. (Sun,) studied this question.