Machine Learning Combining High-Temperature Experiments for the Prediction of Wetting Angle of Mold Fluxes

Abstract

Direct measurement of the wetting angles of the mold fluxes is a strenuous work and time-consuming, and a mathematical model relating the wetting angle of mold flux to its chemical composition is rarely found up to now. In this work, multiple linear regression (MLR), backpropagation neural network (BPNN), and GA-BP neural network (GA-BPNN) are used to model and predict the wetting angle of mold flux. Results show that the accuracy of MLR, BPNN, and GA-BPNN model is 76, 62, and 83 pct; the GA-BPNN model has the highest prediction accuracy. In addition, according to the standardized coefficients in the MLR model, the influence degree of different chemical components on the wetting angle of mold fluxes is analyzed. The importance of the influence of various components on the wetting angle is Fe₂O₃, F⁻, Li₂O, Na₂O, R, MnO, Al₂O₃, B₂O₃, and MgO from high to low. Among them, Fe₂O₃, Li₂O, Na₂O, R, and MnO have a negative effect on the wetting angle of mold flux, while F⁻, Al₂O₃, B₂O₃, and MgO have a positive effect. The established GA-BPNN model could facilitate the design and optimization of mold slag in the steel continuous casting process.