Forecasting secular variation using physics-informed neural networks for IGRF-14.

Abstract

In response to the call for candidate models for the 14th generation of the International Geomagnetic Reference Field (IGRF) by the Geomagnetic Field Modeling Working Group (V-MOD) of the International Association of Geomagnetism and Aeronomy (IAGA), we present the University of Leeds candidate model for the forecast of the average Secular Variation (SV) for 2025-2030. Our approach consists of inverting the geomagnetic field model CHAOS-7.18 using Physics-Informed Neural Networks to produce two global mesh-free models from (i) a mosaic of independent regional flows and (ii) a single global flow representation. The magnetic field is then advected under the assumption of steady core flow over a 5-year period, and the average SV over 5 years is taken to construct the forecast. We validate our approach using hindcasts for the IGRF-13 time period (2020-2025) and benchmark our methodology against the inferred SV from CHAOS-7.18. Our field models constructed from regional flows show reduced RMS misfit relative to the field from the CHAOS-7.18 model at each yearly timestep, compared to the other candidate models from IGRF-13 in the hindcast, both at the Core Mantle Boundary and at the Earth's Surface. We then present our IGRF-14 candidate forecast for the period 2025-2030, derived from the regional method, and discuss possible improvements to this method for future IGRF submissions.