Representation of the Multiphysical Properties of SiO2-Al2O3-CaO Slags by Deep Neural Networks

Abstract

The hybrid analytics and modelling tool is one of the key techniques for the digital twin of industrial processes. Attempt for hybridization of the thermochemical and thermophysical properties of molten SiO2-Al2O3-CaO slags has been recently made at SINTEF Industry. The power of the deep neural networks for the multiphysical modelling of ferroalloy production will be demonstrated in the paper. A deep neural network has been set up for representing thermophysical (electronic conductivity, liquidus temperature, viscosity as well as density) and thermochemical (slag-metal equilibrium composition) properties of the SiO2-Al2O3-CaO system. The physics-based models were first used to model the density, conductivity and viscosity of molten SiO2-Al2O3-CaO slags. Liquidus temperature and metal-slag equilibrium data were calculated using the well-known thermochemical software package, FactSage. These physical models can be used as high-throughput calculation tools for the generation of industrial important labelled data. The DNN model can represent the above labelled metadata with a set of optimized parameters. The DNN model developed in the present study consists of 4 input variables (composition and temperature), 1 hidden layer and 5 output properties was used. The DNN model calculation can almost entirely reproduce all above properties using about 120 parameters. Furthermore, the DNN model can be automatically updated by continuously feeding the experimental data from the laboratory and industry (machine learning). This provides a sound basis for the development of the localized HAM tool for individual ferroalloy producers.