Reconstructing PSP/ISOIS Pitch-angle-resolved Intensities Using Unsupervised Learning

Abstract The Integrated Science Investigation of the Sun (IS⊙IS) on board the Parker Solar Probe provides high-resolution measurements of energetic particles across a broad energy range. Of particular interest are the pitch-angle distributions (PADs), which offer insights into the directionality and transport of solar energetic particles (SEPs). However, 16% of the PAD data from the EPI-Lo instrument are systematically missing during SEP events owing to misalignment between the instrument’s field of view and the local magnetic field direction. These data gaps cannot be reliably addressed using traditional interpolation methods, as PADs exhibit highly structured and nonlinear behavior. In this study, we develop and evaluate a suite of machine learning (ML) models—namely MICE, MISSForest, and M-RNN—to reconstruct missing PAD values across 28 SEP events observed between 2021 May and 2024 March. We find that MICE and a hybrid MICE–M-RNN model consistently outperform interpolation methods and other ML baselines. Approximately half of the SEP events yield high-confidence reconstructions, enabling the partial recovery of uninterrupted PADs and correction of artificial dropouts in omnidirectional intensity timelines. This type of PAD data loss is common across many heliophysics missions with fixed instrument fields of view; therefore, this study proves that ML imputation methods are a good candidate for improving PAD continuity in space-based observations across missions.