Reconstructing Soft X‐ray Photon Count Rates With Missing Segments: A Physics‐Guided Deep Learning Approach

Abstract Soft X‐ray photon count rates are widely used to study interactions between the solar wind and the terrestrial magnetosphere. Soft proton flares frequently contaminate these observations, and screening the contaminated intervals leaves gaps in otherwise continuous time series. We propose a deep learning framework that imputes the missing 0.5–0.7 keV photon count rates from bidirectional temporal context together with simultaneous solar wind and geomagnetic parameters, using the 2.5–5.0 keV band as a background reference. The model employs transformer encoder blocks to leverage bidirectional temporal context and includes a physics‐informed training loss derived from a solar wind charge exchange (SWCX) emission formulation to promote physical consistency. On XMM‐Newton data spanning 2000–2008, the proposed method achieves the lowest median absolute error among all baselines for gaps of 1, 2, 5, and 10 min, supporting reliable downstream analyses in space physics.