Physics-informed neural networks for the improvement of platform magnetometer measurements

Abstract

High-precision space-based measurements of the Earth's magnetic field with a good spatiotemporal coverage are needed to analyze the complex system of our surrounding geomagnetic field. Dedicated magnetic field satellite missions like the Swarm mission form the backbone of research, providing high-precision data with limited coverage. Many satellites carry so-called platform magnetometers that are part of their attitude and orbit control systems. These can be re-calibrated by considering different behaviors of the satellite system, hence reducing their relatively high initial noise originating from their rough calibration. These platform magnetometer data obtained from satellite missions not dedicated to geomagnetic fields complement high-precision data from the Swarm mission by additional coverage in space, time, and magnetic local times. In this work, we present an extension to a previous machine learning approach for automatic in-situ calibration of platform magnetometers. We introduce a new physics-informed layer incorporating the Biot-Savart formula for dipoles that can efficiently correct artificial disturbances due to electric current-induced magnetic fields evoked by the satellite itself. We demonstrate how magnetic dipoles can be co-estimated in a neural network for the calibration of platform magnetometers and thus enhance the machine learning-based approach to follow known physical principles. Here, we describe the derivation and assessment of re-calibrated datasets for two satellite missions, GOCE and GRACE-FO, which are made publicly available. Compared to the reference model, we achieved an average residual of about 7 nT for the GOCE mission and 4 nT for the GRACE-FO mission across all three components combined in the low- and mid-latitudes.