Prediction of titanium burn-off and untimate titanium content in electroslag process

In this study, we investigate the burning behavior of titanium during the electroslag remelting (ESR) process and its impact on the titanium content at the endpoint using machine learning. Initially, a comprehensive database was established by collecting data from literature and experiments, encompassing slag system composition, smelting temperature, and material composition content. Subsequently, six machine learning algorithms, including random forest and Bayesian regression, were employed to model the burning loss behavior of titanium. The random forest model, which exhibited optimal mean square error (MSE) performance, was utilized to generate partial dependence plots. These plots, in conjunction with experimental observations and existing studies, facilitated the analysis of key factors influencing titanium burn loss. Furthermore, the same six machine learning models were applied to predict the endpoint titanium content. The Bayesian regression model demonstrated superior performance in terms of R2 and MSE, leading to the derivation of an empirical formula for predicting endpoint titanium content. This empirical formula was subsequently validated, refined, and optimized using a thermodynamic model based on the theory of molecular ion coexistence. The final prediction formula achieved an error margin of 0.123%.