Science Through Machine Learning: Quantification of Post?Storm Thermospheric Cooling

Machine learning (ML) is often viewed as a black-box regression technique that unable to provide considerable scientific insight. ML models are universal function approximators and - if used correctly can information related the ground-truth dataset for fitting. A benefit over parametric there no predefined basis functions limiting phenomena be modeled. In this work, we develop on three datasets: Space Environment Technologies (SET) High Accuracy Satellite Drag Model (HASDM) density database, spatiotemporally matched of outputs from Jacchia-Bowman 2008 Empirical Thermospheric Density (JB2008), an accelerometer-derived CHAllenging Minisatellite Payload (CHAMP). These compared Naval Research Laboratory Mass Spectrometer Incoherent Scatter radar (NRLMSIS 2.0) model study presence post-storm cooling in middle-thermosphere. We find both NRLMSIS 2.0 JB2008-ML do not account consequently perform poorly periods following strong geomagnetic storms (e.g. 2003 Halloween storms). Conversely, HASDM-ML CHAMP-ML show evidence indicating phenomenon present original datasets. Results reductions up 40% occur 1--3 days depending location strength storm.