Physics-informed Machine Learning for Probabilistic Space Weather Modeling and Forecasting: dB/dt and Geomagnetically Induced Currents

Geomagnetically Induced Currents (GICs) are driven by the interaction of the solar wind and interplanetary magnetic field (IMF). Any long, ground-based conductor is vulnerable to the effects of GICs, including power grids across the globe. Forecasting GICs begins with forecasting dB/dt, or the rate of change of the surface magnetic field. Global magnetohydrodynamic (MHD) models, coupled with other, regional numerical models, are a powerful way to forecast dB/dt distributions and their causes. However, these codes are computationally expensive, preventing broad, statistical investigations of their results. In this work, we will extract and embed into a machine learned model the state-of-the-art knowledge of the underlying physics from simulation output of the Michigan's geospace model, currently operational at the NOAA's Space Weather Prediction Center (SWPC). We will use close to 2 years of operational output data from SWPC. The method will first perform dimensionality reduction for the high-dimensional system using two different approaches and then use a recurrent neural network for nowcast and forecast while accounting for the hysteric properties of the system. The significantly reduced computational cost of the developed model can facilitate both parametric as well as non-parametric uncertainty analysis for probabilistic modeling and forecasting.