Machine Learning and Deep Learning Predictor Models in Estimation of the Thermal State of a Planetary Mantle

Abstract:

We employ synthetic mantle convection models at infinite Prandtl number and 3D-spherical geometry as training and testing samples in training supervised machine learning and deep learning predictive models. In nonlinear systems like the Earth and similar planets, the thermal history of the system is influenced by plate tectonics, the size of core, the radiogenic content of mantle, Rayleigh number, and other physical and rheological parameters. The training samples in our study include numerical convection models with free‐slip and rigid boundary conditions at the surface, as approximations to active plate tectonics and thin surface lithospheres. The predictor models for the surface heat flux and mean mantle temperature include: linear regression predictor with polynomialized features, random forest predictor, and deep learning predictor. The estimation of each predictor model is based on the average estimations obtained from different random choices of training samples (mean field predictor). Our model results reveal prediction accuracies about (97%, 97%), (96%, 98%) and over (99%, 99%) for the mentioned predictor models, respectively, for the surface heat flux and mean mantle temperature.