Designing Structure–Thermodynamics-Informed Artificial Neural Networks for Surface Tension Prediction of Multi-component Molten Slags

The surface tension, as a crucial property of molten slags, affects a broad range of high-temperature industrial processes. In this study, we developed a structure–thermodynamics-informed artificial neural network (STIANN) to predict the surface tension of molten slags over a broad range of composition and temperature. First, we constructed a brand-new database that included not only conventional laboratory-based variable information but also quantitative structural and thermodynamic features at different scales, including second-nearest-neighbor bonds, oxygen species, degree of depolymerization (NBO/T), oxide activities, and Gibbs free energies. Then, the four-layer feed-forward backpropagation artificial neural networks were carefully designed to build the surface tension models. Next, three models were built using the different configurations of training features. The analysis results of structural information indicate the high concentration of bridging oxygen generally contributes to the low surface tension when non-bridging oxygen and free oxygen do the opposite. Statistically, the surface tension is positively correlated with the NBO/T of system. The thermodynamic features of Δ mix G m re and Δ mix G m E vary in the range of 0 to − 70 and 0 to − 55 kJ/mol, respectively, and both decrease first and then increase with the increase in NBO/T. The STIANN model integrated with both structural and thermodynamic information exhibits an unprecedented and excellent predictive performance. The analysis of feature importance confirms the prominent contribution of structural and thermodynamic features to the STIANN model.