High Temperature Melt Viscosity Prediction Model Based on BP Neural Network

This paper comprehensively considers 12 indicators, including temperature, component content, solid–liquid ratio, free volume ratio, atomic cluster as characteristic parameters, to establish a back-propagation (BP) neural network prediction model for the viscosity of multi-element titanium-containing iron-based melts. The comprehensive model is dissected into distinct sub-models based on specific characteristic parameters, including the temperature and composition (T&C)-BP, Liquid structure parameters (LS)-BP, and Solid-phase particle parameters (S)-BP sub-models. The performance and applicability of each sub-model are rigorously analyzed, providing valuable insights into their respective scopes and limitations. By comparing the actual molten iron viscosity with the model predicted value, it was found that, the relative errors for all predicted values were found to be within 10%. The relative error for individual samples at 1350 °C was an impressive 1.3%. Furthermore, a substantial 56% of the predictions exhibited a relative error of less than 5%.