Fig. 1. Framework of the DDM-CNN model for Ms prediction.

Fig. 2. Performance of the SVM models for Wf prediction: (a) the mean results on the training set; (b) the best results on the training set; (c) the mean results on the testing set; (d) the best results on the testing set.

Fig. 3. Distribution of the data in the hierarchical database (n represents the crowding degree): (a) the composition distribution; (b) the composition distribution with a range of 0-5 wt.%; (c) the distribution of the 2nd and 3rd level information.

Fig. 4. Ms prediction results by the DDM-CNN models for the hierarchical database: (a) the mean results on the training set; (b) the mean results on the testing set; (c) the optimal results on the training set; (d) the optimal results on the testing set.

Fig. 5. Relationship between the MAE and crowding degree of the samples in the hierarchical database.

Fig. 6. Comparison of the prediction results by different models on the validation database within the composition range of the original database: (a) the samples in the area of low-alloy steels; (b) the samples in the area of high-alloy steels; (c) the samples beyond the composition range of the original database. (d) Relationship between the MAE and crowding degree of the samples in the validation database.

Fig. 7. Importance of the different elements for ${{W}_{\text{f}}}$.

Fig. 8. (a) Pearson's linear correlation coefficient heat map for different elements. (b) Pearson's linear correlation coefficient of the different models. (c) Comparison of the total error.

Fig. 9. Function of DDM and composition features on Ms prediction.

Fig. 10. Relationship between the MAE and crowding degree of the samples in the validation database for data mining-guided models.

Fig. 11. Performance of the SVM model trained by van Bohemen's database with or without the PAGS information.