Authors: Sanjib K C ( Georgia State University), Viacheslav M Sadykov ( Georgia State University), Dustin Kempton ( Georgia State University)

Radiation measurements at aviation are vital for ensuring the safety of aircraft and crew. Over the past decade, the Automated Radiation Measurements for Aerospace Safety (ARMAS) experiment has collected extensive in-situ data, with over 1,000 flights contributing ~ 400,000 individual measurements across global airspace. Physics-based models—such as the Nowcast of Atmospheric ionizing Radiation for Aviation Safety (NAIRAS)—have been developed to estimate radiation exposure. Nevertheless, it is important to analyze how these models deviate from real-time observations, especially for higher dose rate exposures (>15µSv/hr), and attempt to build full data-driven approaches in order to uncover new relations between radiation environment and Geospace parameters. We investigate the potential of classical linear and nonlinear machine learning (ML) algorithms—including LASSO regression, Random Forest, and XGBoost—for radiation environment nowcasting using a curated ML-ready dataset available from the Radiation Data Portal (https://dmlab.cs.gsu.edu/rdp/). The dataset is partitioned into three distinct subsets for training, validation, and testing. Our analysis focuses on two core objectives: (1) identifying feature importance to compare dominant Geospace properties with those used in physics-based models, and (2) evaluating prediction accuracy using Mean Squared and R-squared error. The results indicate that ML models, particularly XGBoost, slightly but consistently outperform other ML and physics-based models in predictive accuracy across multiple trials. These findings highlight the promise of ML approaches in complementing physics-based modeling techniques, offering new pathways for exploring the global radiation environment and enhancing situational awareness at flight altitudes.