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

• Published in: Space Weather Vol. 20; no. 9
• Main Authors: Licata, Richard J., Mehta, Piyush M., Weimer, Daniel R., Drob, Douglas P., Tobiska, W. Kent, Yoshii, Jean
• Format: Journal Article
• Language: English
• Published: Washington John Wiley & Sons, Inc 01.09.2022
• Subjects: Accelerometers; Aerospace environments; Atmospheric models; Basis functions; CHAMP (experiment); Cooling; Datasets; Density; Drag; geomagnetic storm; Geomagnetic storms; Geomagnetism; Incoherent scatter radar; Laboratories; Machine learning; Magnetic storms; Mass spectrometry; Model accuracy; Modelling; Radar; Satellite drag; Satellites; Scattering; Storms; Thermosphere; Thermospheric cooling; Upper atmosphere
• ISSN: 1542-7390; 1539-4964; 1542-7390
• DOI: 10.1029/2022SW003189

Machine learning (ML) models are universal function approximators and—if used correctly—can summarize the information content of observational data sets in a functional form for scientific and engineering applications. A benefit to ML over parametric models is that there are no a priori assumptions about particular basis functions which can potentially limit the phenomena that can be modeled. In this work, we develop ML models on three data sets: the Space Environment Technologies High Accuracy Satellite Drag Model (HASDM) density database, a spatiotemporally matched data set of outputs from the Jacchia‐Bowman 2008 Empirical Thermospheric Density Model (JB2008), and an accelerometer‐derived density data set from CHAllenging Minisatellite Payload (CHAMP). These ML models are compared to the Naval Research Laboratory Mass Spectrometer and Incoherent Scatter radar (NRLMSIS 2.0) model to study the presence of post‐storm cooling in the middle‐thermosphere. We find that both NRLMSIS 2.0 and JB2008‐ML do not account for post‐storm cooling and consequently perform poorly in periods following strong geomagnetic storms (e.g., the 2003 Halloween storms). Conversely, HASDM‐ML and CHAMP‐ML do show evidence of post‐storm cooling indicating that this phenomenon is present in the original data sets. Results show that density reductions up to 40% can occur 1–3 days post‐storm depending on the location and strength of the storm. Plain Language Summary Machine learning (ML) models are valuable universal function approximators and—if used correctly—can provide scientific information related to the data set used for fitting. A benefit to ML over other common models is that there are no background functions limiting what phenomena it can represent. In this work, we develop ML models on three data sets: the Space Environment Technologies High Accuracy Satellite Drag Model (HASDM) density database, a spatiotemporally matched data set of outputs from the Jacchia‐Bowman 2008 Empirical Thermospheric Density Model (JB2008), and an accelerometer‐derived density data set from CHAllenging Minisatellite Payload (CHAMP). These ML models are compared to the Naval Research Laboratory Mass Spectrometer and Incoherent Scatter radar (NRLMSIS 2.0) model to study the presence of post‐storm cooling in the upper atmosphere. We find that both NRLMSIS 2.0 and JB2008‐ML do not account for post‐storm cooling and perform poorly in periods following strong geomagnetic storms. Conversely, HASDM‐ML and CHAMP‐ML do show evidence of post‐storm cooling indicating that this phenomenon is present in the original data sets. Key Points Machine learning is used to develop models from unique density data sets We compared model predictions along the CHAllenging Minisatellite Payload (CHAMP) orbit during the 2003 Halloween storms We find that models developed on CHAMP and High Accuracy Satellite Drag Model density data can capture density depletion in the post‐storm period