Multi‐Model Ensembles for Upper Atmosphere Models

Abstract

Introduction BackgroundAccurately propagating satellite orbits requires knowledge of the forces acting on the satellite. For satellites in low Earth orbit (less than 1,000 km), forces include terrestrial gravity, solar radiation pressure, lunar and solar gravity and drag caused by the atmosphere (Eshagh & Najafi Alamdari, 2007). The drag force increases dramatically as a satellite's altitude decreases and becomes significant below approximately 600 km (Fortescue et al., 2011). However, there are large uncertainties in modeling the magnitude of the drag acting on a satellite. To do so requires an understanding of the thermospheric mass density, winds and the satellite's ballistic coefficient. The largest contribution to error in the forecasting of satellite positions is specification of thermospheric density (Mehta et al., 2018), although for tumbling or complex geometries, the errors in the ballistic coefficient can be a substantial contribution.Currently a variety of mathematical models are used to provide estimates of the density. Empirical models are often used by satellite operators. They are fitted to measurements of thermospheric parameters; however such measurements are sparse. In particular, there are very few measurements between 100 and 250 km because balloons cannot reach these heights and satellites re-enter too quickly for any long term study. Fabry-Perot Interferometers can be used to measure wind between 220 and 600 km (Titheridge, 1995) and meteor radars can measure wind, as well as temperature and pressure, between 80 and 100 km (John et al., 2011;Reid et al., 2018).Physics-based models solve the equations which describe the physical processes in the thermosphere. Initially the atmospheric density, wind and temperatures are generally provided by empirical models, but a "spin-up" time is used for the results to stabilize. The spin-up time can be reduced in subsequent model runs by using previous output from the model. Neutral and ion species production is then calculated via chemical reaction equations and using solar X-rays and EUV conditions. Ion transportation and recombination are also considered. The initial and boundary conditions, as well as proxies for solar activity, are the main drivers for the models. There are a number of approaches to modeling the physics of the thermosphere, which rely on different numerical methods

Projected heatwaves in Xinjiang Uygur autonomous region, China

“…MME is an effective method to reduce uncertainty (Parsons et al, 2021). The critical consideration in utilizing MME is how to combine models, various weighting schemes include equal weighting, performance weighting, performance weighting with bias removed, reliability ensemble averaging, independent weighting, and non-negative least squares regression (Elvidge et al, 2023). In this study, to enhance the reliability of projections, we evaluate the simulation capabilities of individual models in terms of interannual variability and spatial patterns are assessed.…”

Statistical models of the variability of plasma in the topside ionosphere: 2. Performance assessment

“…Bias in physical models is relatively common (i.e. Elvidge et al, 2023), raising the possibility that the Swarm-VIP models could be used to calibrate the physical models. This possibility will be explored in a subsequent paper which will also conduct a more detailed comparison of TIE-GCM and the Swam-VIP models.…”

Probabilistic Short‐Term Solar Driver Forecasting With Neural Network Ensembles

“…Although an improvement was seen, one must consider that certain models are more skilled than others, well performing models should be weighted more heavily than models with worse performance. Weighted ensembles performed better than unweighted for upper atmosphere models (Elvidge et al, 2023). We consider a stacked ensemble approach (Sridhar et al, 1996), which uses linear regression to optimally combine predictions.…”