Can deep learning help in understanding the geodynamo?

Abstract

Machine learning (ML) is an emerging tool used in an ever-growing number of branches of contemporary geophysical research. To date however, rarely have such methods been applied specifically for the study of Earth’s internal magnetic field and the geodynamo. Prevailing methods currently used in inferring the characteristic properties and the probable time evolution of the geodynamo are based on two differing but connected approaches. They both directly rely on simplified applications of magnetohydrodynamics. In this study, first attempts at applying a new kind of inference method are presented. Its long-term goal focuses on inferring concentrations of electric current densities inside the core as the direct sources of the magnetic field itself rather than addressing the fluid motions. The method makes use of simplified models in which equivalent current loops represent local current systems emerging in computationally extremely intensive turbulent geodynamo simulations. These forward models were then utilized to produce synthetic magnetic field maps with a widely varying configurations of such current loops hypothetically placed inside a conductive outer core, along maps of real historical geomagnetic field model data both computed at the core-mantle boundary (CMB). The resulting maps were used as input training samples for an image processing deep neural network, which essentially learns to infer the parameters of all these modelled current systems based on the magnetic field maps. Results presented here are focusing mainly on the feasibility of the method when applied for synthetic data and the quality parameters of these inferences. The overall distribution of loop currents can be recovered with a satisfying accuracy. However, the quality of the reconstruction of magnetic field properties is compromised by unrealistic values occasionally being inferred, and an inability to recover realistic secular time variations (SV). One example is shown of an inferred model obtained using CMB magnetic field maps. Main features of the maps of actual field values can be reproduced, and the corresponding inferred current loops tend to align in chain-like arrangements.