Ground magnetic field perturbation forecasting based on deep learning on spherical harmonics decomposition

Ground magnetic field perturbation is currently estimated based on Magnetic Hydrodynamic (MHD) and empirical models. While useful, MHD models are computationally expensive for high-resolution models that are required for small-scale perturbation, and empirical models do not provide a dynamic forecast. In this study, we used Spherical Harmonics (SH) to create high-resolution global models of dBH and dBH/dt with predictive lead time using ML techniques. We apply a compressed sensing technique to decompose global geomagnetic variations in terms of spherical harmonics. Our method allows for reconstruction of the global perturbations down to 1 min cadence (the available cadence from the SuperMAG network). We proceed to train Recurrent Neural Networks (RNNs) to forecast the evolution of spherical harmonic coefficients in response to solar wind conditions measured at L1 point (OMNI). Evaluation metrics such as the R2 score, mean absolute error, and Heidke Skill Score (HSS) were used across all stations for both the reconstruction and prediction model. We compare our result with state of the art Wiemer model as well as developed forecasting models in the literature using the HSS score.