Development of structure-informed artificial neural network for accurately modeling viscosity of multicomponent molten slags

The design and optimization of many high-temperature industrial processes have great demand for viscosity models of molten slags. Due to the unsatisfactory performance of conventional models, we developed a structure informed artificial neural network (SIANN) model for the first time to predict the viscosity of molten slags. The model database containing 1892 measurement values was constructed from carefully identified literature and covered the temperature, compositional, and structural spaces. The feed-forward four-layer perceptron artificial neural network was designed to capture the complex dependence of viscosity upon influence factors (composition, temperature, and structure). The result indicates that after quantitative atom-level information is integrated into the model, its ability to accurately predict viscosity gets significantly improved. The interpretability of the obtained SIANN mode is highlighted with selected structural features that have a strong determinant on viscosity. Furthermore, the comparisons of prediction performance indicate the obtained model outperforms other existing models, achieving the minimum predicted deviation in various component systems.

Automated summary

A structure informed artificial neural network (SIANN) model was developed for predicting the viscosity of molten slags, integrating quantitative atom-level information to significantly improve prediction accuracy. The interpretability of the model was highlighted by selected structural features with strong determinant impact on viscosity. The obtained model outperformed other existing models in terms of prediction performance across various component systems.