Abstract:
Viscosity is one of the important factors in understanding the Earth’s thermo-chemical evolution from the magma ocean stage, as well as in understanding processes such as lava eruption, extent of lava flows on the surface. Since viscosity controls the eruption method, transport dynamics of melts, different melts at different temperatures, pressure, and composition show different settling and fractional crystallization characteristics and thus can be helpful in modeling earth processes right from the magma ocean stage to partial melting of the mantle and crust formation.
It is always not possible to determine viscosity of silicate melts of so many different chemical compositions experimentally at all the extreme temperature and pressure conditions present in the Earth’s interior or during volcanic eruptions. We have developed machine-learning models to predict magma viscosity for a range of silicate melt compositions at high temperatures (4000C – 15000C).
We have compiled experimentally determined viscosity data from various research publications to generate our master database to train our model. Data consists of log of viscosity (Pa-s) values against temperature© and weight percentage composition of volatiles (Water, Fluorine, Loss on ignition) along with weight percentage compositions of minor and major oxides (SiO2, Al2O3, TiO2, FeO, Fe2O3, CaO, MgO, MnO, Na2O, K2O, P2O5). We have compiled 2294 data points out of which 1792 are completely anhydrous compositions. In this project, three models 1) linear regression, 2) Random forest, and 3) Xg-boost are developed. Xg boost model outperforms all models giving out the lowest error in predicting magma viscosity for a given composition under high temperature conditions. Additionally, we have also discussed the relative importance of different variables affecting the viscosity of silicate melts
