AbstractA data‐driven model of Earth's magnetosheath is developed by training a recurrent neural network (RNN) with probabilistic outputs to reproduce Magnetospheric MultiScale (MMS) measurements of the magnetosheath plasma and magnetic field using measurements from the Wind spacecraft upstream of Earth at the first Earth‐Sun Lagrange point (L1). This model, called Probabilistic Regressor for Input to the Magnetosphere Estimation‐magnetosheath (PRIME‐SH) in reference to its progenitor algorithm PRIME, is shown to predict spacecraft observations of magnetosheath conditions accurately in a statistical sense with a continuous rank probability score of 0.227σ (dimensionless standard deviation units). PRIME‐SH is shown to be more accurate than many current analytical models of the magnetosheath. Furthermore, PRIME‐SH is shown to reproduce physics not explicitly enforced during training, such as field line draping, the dayside plasma depletion layer, the magnetosheath flow stagnation point, and the Rankine‐Hugoniot MHD shock jump conditions. PRIME‐SH has the additional benefits of being computationally inexpensive relative to global MHD simulations, being capable of reproducing difficult‐to‐model physics such as temperature anisotropy, and being capable of reliably estimating its own uncertainty to within 3.5%.
Read full abstract