Soft Proton Environment Modeling for SMILE Using XGBoost and Solar Cycle–Informed Synthetic Data Sets

Abstract Soft proton contamination presents a persistent threat to the performance of space‐based X‐ray observatories such as ESA's upcoming Solar‐wind Magnetosphere Ionosphere Link Explorer (SMILE). This study develops a machine‐learning framework for modeling soft proton fluxes in the 92.2–159.7 keV range, tailored to SMILE's elliptical, highly inclined orbit. Using past data from ESA's Cluster mission and a solar‐cycle‐based extrapolation scheme derived from Solar Cycle 23, we simulate the average space weather conditions SMILE is likely to encounter during its first 3 years of operation. The proposed model leverages XGBoost, a nonlinear ensemble learning algorithm, and introduces a rotating quarterly train–test split strategy to overcome distributional mismatches present in earlier models. This new approach improves generalization and substantially reduces underprediction bias during periods of elevated proton flux, achieving a test set of 0.45 and reducing bias from −0.30 to −0.10 log(1/c/s/sr/keV). Analysis of the spatial distribution of the models' predicted proton flux for the SMILE mission shows good agreement with known magnetospheric structures, including lobe/cusp regions and solar‐cycle‐dependent enhancements. The results of this study provide a data‐driven estimate of the ambient soft proton environment along the SMILE orbit.