Improving Platform Magnetometer Measurements Using Physics-informed Neural Networks

So-called platform magnetometers, mounted on a variety of non-dedicated satellites in Low-Earth orbit, are promising instruments to increase the spatiotemporal coverage of space-based measurements of the Earth’s magnetic field. However, these instruments often need to be calibrated to ensure a scientific accuracy and usability of the data they collect. To do this, it is important to gather information about the satellite to correct artificial disturbances caused by other payload systems as well as other influencing properties. In the past, we demonstrated that a Machine Learning-based calibration achieves competitive results. By using machine learning techniques, the magnetometer signal can be adapted to account for artificial disturbances and the proposed non-linear regression method can automatically identify relevant features and their crosstalk, enabling the use of a wider range of inputs. This reduces the analytical work required for the calibration of platform magnetometers, resulting in faster, more precise, and easily accessible magnetic datasets from non-dedicated missions. The calibrated datasets are made publicly available.

In this work, we propose an extension for the known approach by incorporating the physical Biot-Savart formula into a neural network, which results in a physics-informed neural network. This improves the modeling and correction of the impact of current-induced artificial magnetic fields on the satellite and its magnetic measurements. In addition, the Average Magnetic field and Polar current System (AMPS) model is combined with the CHAOS-7 model, improving the reference model of the calibration, especially for the polar regions. This extended approach is applied to the GOCE and GRACE-FO satellite missions and their respective measurements. In the future, the underlying software shall be published and applied to a wider variety of satellites to improve the accuracy of their platform magnetometer measurements. By making this tool publicly available, we hope to enable other satellite operators to calibrate their instruments, improve the quality of their data, and make additional data available to the scientific community.