Modeling Viscosity of Volcanic Melts With Artificial Neural Networks

Abstract Viscosity is of great importance in governing the dynamics of volcanoes, including their eruptive style. The viscosity of a volcanic melt is dominated by temperature and chemical composition, both oxides and water content. The changes in melt structure resulting from the interactions between the various chemical components are complex, and the construction of a physical viscosity model that depends on composition has not yet been achieved. We therefore train an artificial neural network (ANN) on a large database of measured compositions, including water, and viscosities that spans virtually the entire chemical space of terrestrial magmas, as well as some technical and extra-terrestrial silicate melts. The ANN uses composition, temperature, a structural parameter reflecting melt polymerization and the alkaline ratio as input parameters. It successfully reproduces and predicts measurements in the database with significantly higher accuracy than previous global models for volcanic melt viscosities. Viscosity measurements are restricted to low and high viscosity ranges, which exclude typical eruptive temperatures. Without training data at such conditions, the ANN cannot reliably predict viscosities for this important temperature range. To overcome this limitation, we use the ANN to create synthetic viscosity data in the high and low viscosity range and fit these points using a physically motivated, temperature-dependent viscosity model. Our study introduces a synthetic data approach for the creation of a physically motivated model predicting volcanic melt viscosities based on ANNs.