Learning Thermospheric State Evolution: An Adaptive Neural Operator Framework Based on TIE-GCM Simulations

Reliable short-term prediction of thermospheric states is important for satellite drag applications but remains difficult because of nonlinear, multiscale variability. We developed a multivariable Adaptive Fourier Neural Operator (AFNO) surrogate using 24 years (2000–2023) of Thermosphere–Ionosphere Electrodynamics General Circulation Model (TIE-GCM) simulations. The model predicts neutral density, temperature, winds, and geopotential height over 24-h autoregressive forecasts. We compared direct state prediction with increment-based flow prediction and tested logarithmic density scaling and one-hour historical inputs. Both formulations preserved dominant large-scale density structures and the equatorial mass density anomaly, with anomaly correlation coefficients above 0.94 for all variables. The Direct formulation maintained lower errors and greater stability at longer lead times, whereas Flow performed better only at early steps. Logarithmic density scaling produced variable- and altitude-dependent trade-offs, and historical inputs yielded no overall benefit. During a representative geomagnetic storm, the baseline reproduced broad density morphology but increasingly underestimated enhancement magnitude with lead time. Because evaluation used the independent 2023 TIE-GCM test year with prescribed forecast-time forcing and no observational validation, the framework should be interpreted as a TIE-GCM-consistent surrogate rather than a validated predictor of the observed thermosphere.