Deep-Learning-Based Broadband Lightsource X-ray Absorption Spectroscopy Using Photon-Counting Detector

Photon-counting detectors enable high-resolution, multienergy X-ray spectroscopy but suffer from substantial spectral distortions under high photon flux due to pile-up and polarization effects. We propose Spectral-Transformer, a deep learning framework for correcting these distortions in broadband lightsource X-ray absorption spectroscopy (BL-XAS). The model fuses spectra and tube current inputs through a bimodal mapping mechanism and applies physics-informed loss functions as feedback to preserve peak positions and total photon numbers. To facilitate robust model training, we also developed a cascade simulation framework that generates high-fidelity training data by coupling the nonlinear effects of polarization and pile-up. Experimental results validate our approach, showing that the corrected spectra achieved exceptional fidelity. Compared to the raw distorted spectra, the Spectral-Transformer reduces the KL divergence to 1.0 × 10⁻⁴. This enhanced spectral integrity directly translates to superior analytical performance, boosting material classification accuracy from 86.2% to 95.5%.