Melting temperature prediction method for mold flux based on transfer learning algorithm

To establish an accurate predictive model for the single hot thermocouple technique (SHTT)-measured melting temperature, a transfer learning strategy was introduced. The proposed model learns general relationships between mold flux composition and melting behavior on the large hemisphere-method dataset in the pre-training process and retains them in the feature-extraction layers. Fine-tuning the output layer with limited SHTT data then adapts the model to SHTT measurements, alleviating the accuracy bottleneck associated with direct training on a small SHTT dataset. The results show that the coefficient of determination ( R 2 ) and mean squared error (MSE) of the pre-trained model (Hemi_MLP) on the test set are 0.87 and 1,268.65, respectively. It provides effective general knowledge of mold flux melting behavior for the subsequent fine-tuning task. The fine-tuned model (SHTT_MLP) achieves 0.99 R 2 and 187.6 MSE on the SHTT test set, which has the best performance among the comparison models. The Shapley Additive Explanations analysis shows that the determination of the SHTT_MLP for the influence of each component on the melting temperature of mold flux is consistent with prior knowledge. Additionally, the SHTT_MLP primarily completes the fine-tuning process by strengthening its sensitivity to CaO/SiO 2 and CaO/Al 2 O 3 .