SILICON AND OXYGEN IN EARTH’S CORE: APPLICATIONS OF MACHINE LEARNING ALGORITHMS TO METAL-SILICATE EQUILIBRIA AND CORE FORMATION

Within Earth’s core, light elements (Si, O, C, S, N, H) are known to make up a small fraction of the total mass of the core with respect to heavy elements. The degree to which these elements exist in the cores of terrestrial planets have geophysical and geochemical implications, most notably the presence of core convection and a geodynamo, thermal conductivity within the core, and core temperature. Comparison of the composition of chondrites to Earth’s mantle composition and the Preliminary Reference Earth Model have given an estimation of about 10% light elements in Earth’s core. The concentrations of each light element have been characterized in previous literature by determining experimentally the partitioning of elements between metal and silicate phases at high pressure and temperature. Previous studies have constructed thermodynamic models using linear regressions, to predict the change of partition coefficients with pressure, temperature, and oxygen fugacity. However, there is a large variance among previous literature in resulting thermodynamic models, which is likely indicative in substantial regression errors. Here, we used machine learning algorithms, including Random Forest Regressions and Neural Networks, to predict the partition coefficients of Si and O using MetSilDB, a database for metal-silicate equilibria (Boujibar et al., GSA Fall 2022 Conference). We assessed the accuracy of our models using cross-validation techniques. Using these methods, we built a model predicting elemental partitioning coefficients with a highly improved performance than previous models. In addition, machine learning algorithms enabled addressing non-linear effects of experimental variables. Our findings will help us infer the composition of the core more accurately.