Fig. 1. A flowchart of the ML-based design strategies in low-alloy steel. Strategy I: an ML model with composition ${{C}_{i}}$ as input features; Strategy II: an ML model with composition ${{C}_{i}}$ and heat treatment parameters ${{H}_{i}}$ as the input features; Strategy III: an ML model with composition ${{C}_{i}}$, heat treatment parameters Hi, and physical features ${{P}_{i}}$ as input features.

Fig. 2. Correlation matrix for the original 42 features and the targeted CIT. Positive and negative correlations were indicated by blue and red, respectively. In the upper-right part, the greater the filling score of the circles in each pie chart, the better the correlation. In the lower-left part, circles with larger areas denote higher relevance.

Fig. 3. Further selection of facile features for the CIT model. (a) PCC and RMSE of each feature in a particular model, where features of the same cluster were marked with the same color. (b) The importance sequence of features to the CIT is evaluated by FSelector. (c) PCC and (d) RMSE of each possible linear model contain a subset of the six features in our training data. The red font indicates the best performers.

Fig. 4. Scatter plot matrices showing the correlation between TT, $\text{dA}{{\text{R}}^{\text{Fe}}}$, TS, RA, and CIT in the studied low-alloy steel. The color represents the value of CIT.

Fig. 5. Experimental values vs the predicted values by the optimal ML models that used (a)${{C}_{i}}$, (b) ${{C}_{i}}$ and ${{H}_{i}}$, (c)${{C}_{i}}$, TT, $\text{dA}{{\text{R}}^{\text{Fe}}}$, TS, and RA as features. Note: all data points align along the 45° diagonal line indicating a perfect model. (d) PCC and MAE of six typical ML models. (All models are evaluated by 10-fold cross-validation).

Fig. 6. Experimental values vs predicted values by the formula. Note: all data points align along the 45° diagonal line indicating a perfect model. (The symbolic regression process was evaluated by the 10-fold cross-validation).