Machine learning-based predictions of source, seed, and relativistic electron phase space density in the outer radiation belt.

The article focuses on the development of an artificial neural network (ANN) model to predict electron phase space density (PSD) in Earth's radiation belts using in situ measurements from the Van Allen Probes (2012–2019) combined with solar wind and geomagnetic indices. Covering 31 energy channels (μ = 10¹–10⁴ MeV/G) and two values of the second adiabatic invariant (K = 0.08 and 0.17 G¹/²R_E) across L* shells from 2.0 to 5.5, the model achieves high accuracy in the outer radiation belt core (L* = 3.0–5.5), particularly for source, seed, and relativistic electrons. Performance metrics include root mean square errors below 0.155, prediction efficiencies above 0.982, and Pearson correlation coefficients exceeding 0.981, with over 90% of predictions within 0.2 orders of magnitude of observations. The model also effectively reproduces the energy spectral evolution of electron PSD during geomagnetic storms, offering a valuable tool for advancing space weather forecasting and mitigating risks to satellite operations.