Nowcasting of Aviation Radiation Using Geospace Environment Properties: A Machine Learning Approach

TL;DR

The study addresses real-time nowcasting of aviation radiation by leveraging a machine learning pipeline trained on the Radiation Data Portal dataset, using 47 feature inputs plus ARMAS dosimetry targets. It compares LASSO, Random Forest, and XGBoost against the NAIRAS-v3 physics baseline, finding that XGBoost delivers the best predictive performance with an average RMSE around 3.67 μSv/h and $R^2$ near 0.54, outperforming NAIRAS-v3 (RMSE ≈ 4.05 μSv/h). Feature-importance analyses reveal that geomagnetic cutoff rigidity, GPS altitude, solar polar fields, and solar wind properties are key drivers of aviation radiation nowcasting. The results support the viability of ML-based approaches for operational space weather applications and motivate future work on time-series models to capture temporal dynamics in the Geospace environment.

Abstract

Radiation exposure at aviation altitudes presents significant health risks to aircrews due to the cumulative effects of ionizing radiation. Physics-based models estimate radiation levels based on geophysical and atmospheric parameters, but often struggle to capture the highly dynamic and complex nature of the radiation environment, limiting their real-time predictive capabilities. To address this challenge, we investigate machine learning (ML) methods to enhance real-time radiation nowcasting. Leveraging newly compiled ML-ready datasets, publicly available at https://dmlab.cs.gsu.edu/rdp/, we train supervised models capable of capturing both linear and non-linear relationships between Geospace conditions and atmospheric radiation levels. Our experiments demonstrate that the XGBoost model achieves approximately 10 percent improvement in prediction accuracy over the considered physics-based model. Furthermore, feature importance analysis reveals that certain Geospace properties, specifically solar polar fields, solar wind properties, and neutron monitor data, are impacting the nowcast of the radiation levels at flight altitudes. These findings suggest meaningful physical relationships between the near-Earth space environment and atmospheric radiation, and highlight the potential of ML-based approaches for operational space weather applications.

Figures (6)

• Figure 1: Effective dose rate without shielding along the Tokyo–Atlanta flight trajectory on June 7, 2025, calculated from NAIRAS simulation using the Run-on-Request (RoR) services of the Community Coordinated Modeling Center.
• Figure 2: Comparison of the Root Mean Squared Error (RMSE) in radiation prediction for different train-calibration-test partition combinations. Note: '123' means partition 1 is train, partition 2 is validation, and partition 3 is test, and likewise.
• Figure 3: Measured radiation dose rates VS NAIRAS-v3 (top) and nowcasted XGBoost model (bottom) for partition 1 (split-231). Both the NAIRAS-v3 and XGB nowcasts demonstrate the poor performance for the ‘tail’ of the distribution (measurements above $\sim$15 µSv/h).
• Figure 4: Feature importance according to the LASSO model. The vertical lines show 1%, 5%, 10%, and 25% normalized importance.
• Figure 5: Mean feature importance according to the Random Forest model.