Abstract: The increasing demand for high-performance steels and environmentally sustainable pyrometallurgical processes has led to significant compositional variations in ionic melts and the development of novel fluxes. Optimizing ionic melt performance requires precise control of thermophysical properties, with viscosity being a key factor influencing heat and mass transport in various industrial applications. However, traditional experimental and analytical methods are often cost-prohibitive and pose challenges in generalizing findings across diverse compositions and temperatures. This study introduces MOVINet (MOlten ions VIscosity Network), a data-driven modeling framework designed to predict high-temperature ionic melt viscosity based on melt composition, temperature, and fundamental properties of 13 components, including 12 oxides and one fluoride. Trained on 1981 experimentally measured data points and evaluated using 480 independent data points, MOVINet achieved a mean absolute error (MAE) of 0.1480, reducing error by 57.7 % compared to the best existing model (MAE = 0.3497). It consistently demonstrated high accuracy across six ionic melt types over a broad temperature range (1100-1870 °C) and maintained low errors even for melts containing previously unseen components (e.g., MAEs of 0.0567 for CaCl₂ and 0.1463 for BaO-containing samples). Furthermore, explainable AI analysis confirmed the dominant influence of temperature while highlighting compositional features affecting viscosity.

Key words: Viscosity, Ionic melts, Neural network, Data-driven approach, Explainable AI