Thermal conductivity of Fe-bearing bridgmanite and post-perovskite determined by machine learning

Bridgmanite (Brg) and post-perovskite (PPv), as the most abundant minerals at the lower mantle, could be the main minerals in the Large Low Shear Velocity Provinces (LLSVPs). However, their thermal conductivities, as well as the impact of Fe impurities, are highly controversial. Measuring the thermal conductivity of minerals at high P-T conditions remains challenging, and determining the thermal conductivity of minerals by first principles calculations leads to finite size effects due to computational limitations. To overcome computational limitations, we trained a machine learning potential for Fe-free and Fe-bearing Brg and PPv with data from first-principles calculations, then investigated their thermal conductivity at high P-T conditions based on the machine learning potential in the large cells with finite-size effects well considered. We found that the presence of 12.5 mol% Fe in the lowermost mantle decreases the thermal conductivities of Brg and PPv by 10% and 14%, respectively. Furthermore, the phase transition from Brg to PPv increases the thermal conductivity of pyrolite by 22%. Incorporating the distribution of minerals, temperature, and iron content obtained through the inversion based on mineral elasticity and seismic tomography models, we found the heat flux is significantly lower in the LLSVPs regions, which would have important implications for the geomagnetic field and the thermal evolution of the Earth.