Deep learning-based electronics architecture for pulse pile-up correction in synchrotron radiation spectroscopic detection

Synchrotron light sources have become indispensable tools for microstructural analysis, material characterization, and dynamic process observation, owing to their high photon flux and broad wavelength range. However, the increased photon flux introduces challenges in maintaining energy resolution in spectroscopic detection, particularly due to pulse pile-up at high count rates. This paper proposes a novel hybrid architecture that integrates deep learning algorithms with FPGA hardware to achieve real-time pulse pile-up correction in synchrotron radiation spectroscopic detectors. By employing the Transformer algorithm within an FPGA and leveraging GPU-based retraining to adapt to varying count rates, the system dynamically optimizes energy resolution. Experimental validation with data from the VortexEX90 detector demonstrates that the proposed approach achieves an energy resolution of approximately 126.8 eV at 60 kcps and 135.5 eV at 1 Mcps for incident photons of 22 keV, ensuring good performance across a wide range of count rates. This work provides a scalable, high-efficiency solution for pulse pile-up correction, advancing the capabilities of synchrotron spectroscopic detection under high-count-rate conditions.