Data-driven machine learning methods to predict the viscosity-temperature characteristics for gasification slags

The viscosity of molten silicate slags is crucial for the stable operation of the gasifier. However, there is currently no effective prediction method for the viscosity of non-Newtonian slags because the complex components lead to the diversification of crystals, making it difficult to accurately predict the viscosity of non-Newtonian slags based on crystallization or flow mechanisms. In this study, advanced artificial intelligence methods were used to predict the viscosity of coal slags through data-driven approaches. Four machine learning models were established—Random Forest (RF), Gradient Boosting Decision Tree (GBDT), Backpropagation (BP) Neural Network, and Support Vector Regression (SVR)—with a collection over 2,420 experimental data points as the dataset. Before constructing the models, an analysis of the chemical properties of gasification slags is conducted to identify the factors that significantly influence viscosity. Then the compositions Si, Al, Ca, Fe, Mg, Na, K, Si/Al as well as temperature are selected as the input parameters. After training and test, results show that viscosity predication based on classified dataset on fluid types performs better than that without classification. Among the four models, the BP Neural Network model shows the best with least error for both Newtonian and non-Newtonian slags. Besides, the Critical Viscosity Temperature (T_cv) is demonstrated with a good prediction by BP Neural Network model.