Decoding viscosity-microstructure relationships in the ternary CaO-SiO2-FexO system via integrated machine learning and multimodal characterization

Abstract

This study investigates the viscosity and microstructure of the ternary CaO-SiO₂-Fe_xO system using a combination of deep neural network (DNN) learning, in-situ Raman spectroscopy, X-ray photoelectron spectroscopy (XPS), and quantum chemical ab initio calculation. A DNN-based viscosity prediction model was developed using a dataset of 1483 experimental data points, which were partitioned into training, validation, and test sets at a 5:2:3 ratio for model training and evaluation. The model achieved high prediction accuracy with a coefficient of determination (R²) of 0.9464 and a mean absolute error (MAE) of 0.069. The dataset encompasses the key compositional range of metallurgical slags, spanning 0–70 mol% SiO₂, 0–70 mol% CaO, and 0–90 mol% Fe_xO. The model enables rapid and accurate viscosity predictions, reducing the need for extensive experimental measurements. Microstructure analysis via XPS and Raman spectroscopy revealed that with increasing iron content, the silicon-oxygen tetrahedron (SiOT) network structure is disrupted, leading to a transformation from Si–O–Si to Si–O–Fe and Fe–O–Fe bonds, accompanied by a decrease in viscosity. This study also quantitatively correlates the Fe³⁺/(Fe³⁺+Si⁴⁺) ratio in tetrahedral coordination with melt viscosity through structure descriptors (NBO/Si ratio). These results demonstrate that an increase in the tetrahedral Fe³⁺/(Fe³⁺+Si⁴⁺) ratio nonlinearly elevates NBO/Si values (correlation coefficient r = 0.99), which linearly reduces melt viscosity (r = 0.96) through depolymerization of the SiOT network. The established model provides a predictive framework for viscosity optimization in metallurgical slag design and quantitative analysis of magma transport dynamics in geological systems.