Machine learning (ML) as a tool is rapidly emerging in various branches of contemporary
geophysical research. To date, however, rarely has it 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 mostly based on reduced representations of magnetohydrodynamics
(MHD). This study introduces a new inference method, referred to as Current Loop-based
UNet Model Segmentation Inference (CLUMSI). Its long-term goal focuses on uncovering
concentrations of electric current densities inside the core as the direct sources
of the magnetic field itself, rather than computing the fluid motion using MHD. CLUMSI
relies on simplified models in which equivalent current loops represent electric current
systems emerging in turbulent geodynamo simulations. Various configurations of such
loop models are utilized to produce synthetic magnetic field and secular variation
(SV) maps computed at the core–mantle boundary (CMB). The resulting maps are then
presented as training samples to an image-processing neural network designed specifically
for solving image segmentation problems. This network essentially learns to infer
the parameters and configuration of the loops in each model based on the corresponding
CMB maps. In addition, with the help of the Domain Adversarial Training of Neural
Networks (DANN) method during training, historical geomagnetic field data could also
be considered alongside the synthetic samples. This implementation can increase the
likelihood that a network trained primarily on synthetic data will appropriately handle
real inputs. Our results focus mainly on the method's feasibility when applied to
synthetic data and the quality of these inferences. A single evaluation of the trained
network can recover the overall distribution of loop parameters with reasonable accuracy.
To better represent conditions in the outer core, the study also proposes a computationally
feasible process to account for magnetic diffusion and the corresponding induced currents
in the loop models. However, the quality of the reconstruction of magnetic field properties
is compromised by occasional poor inferences, and an inability to recover realistic
SV.
