Code and Storm-Event Data for Robust Short-term Forecasting of Thermospheric Density and Orbit Prediction Using an Interpretable iTransformer Model

Reproducibility materials for the manuscript “Robust Short-term Forecasting of Thermospheric Density and Orbit Prediction Using an Interpretable iTransformer Model.” This archive contains the Python implementation of the iTransformer-based thermospheric-density forecasting model and the residual-based Bi-LSTM baseline, together with the trained iTransformer checkpoint, normalization parameters, training configuration and records, and storm-event data used in the manuscript. The iTransformer model uses the preceding 24 hours of thermospheric-density observations, together with spatiotemporal and space-weather covariates, to forecast density over the following 24 hours. The implementation includes variate tokenization, Reversible Instance Normalization (RevIN), AdamW optimization, a three-epoch learning-rate warm-up followed by cosine decay, gradient-norm clipping, early stopping, and checkpoint selection based on the minimum validation MSE. The archived storm-event datasets contain observations and model predictions for the March 2008 CHAMP and March 2015 GRACE-A cases and were used for the event-based analyses reported in the manuscript, including the results summarized in Table 6. The archive also provides the trained iTransformer model checkpoint and associated normalization parameters to facilitate model evaluation and reuse. External datasets required for complete model retraining should be obtained from their original repositories, as described in the manuscript.