Prediction of nucleation lag time from elemental composition and temperature for iron and steelmaking slags using deep neural networks

Abstract

A prediction of the nucleation lag time of iron and steelmaking melts solely from elemental composition and temperature was produced via deep neural networks trained on data available in the literature. To the best of our knowledge, this constitutes the first published instance of prediction of nucleation lag time that does not require composition specific empirical data. Control of the nucleation process is critical for the production of ground granulated blast furnace slag, control of slag properties for heat recovery or utilization, and the optimization of slag for CO2 mineralization. The deep neural network achieved an average absolute scaled error (AASE) over a testing set of 947 points covering 7 orders of magnitude of 39.9%. Performance was further improved by bootstrapping with a prediction of liquidus temperature from a separate deep neural network (AASE = 33.4%). Bootstrapping using DNN-generated viscosity data did not increase prediction accuracy. The negligible calculation load of the trained deep neural networks allows for rapid design, analysis, and optimization of novel slag compositions and treatment methods. This ability was demonstrated by calculating the necessary continuous cooling rate to generate amorphous slag across all CaO–Al2O3–SiO2 and CaO–FeO–SiO2 compositions and the potential to use additives to alter said cooling rate.