Transformer Algorithm for Pile-Up Correction in Synchrotron Radiation Spectroscopic Detection Experiments

In this study, a Transformer-based algorithm is proposed for pulse pile-up correction in synchrotron radiation spectroscopy detection experiments. The Transformer model, employed as a deep learning framework, is evaluated in conjunction with Gaussian and trapezoidal filter shaping. Through systematic optimization of the model architecture and hyperparameters, its impact on recognition accuracy, loss rate, and generalization performance is analyzed. Additionally, the influence of varying the number of sublayers within the Transformer is investigated, showing that an increased number of sublayers enhances the model’s capacity to process complex signals and improves generalization ability. Test experimental results indicate that trapezoidal filter shaping outperforms Gaussian filtering, demonstrating faster convergence and higher recognition accuracy, particularly when addressing pile-up signals. Spectroscopic experiments on iron, cobalt, and copper further demonstrate that the Transformer model consistently maintains energy resolution within the 125-136 eV range, even under high count rate conditions, supporting its applicability to high-throughput synchrotron environments. These findings highlight the potential of the Transformer algorithm to enhance real-time pulse pile-up corrections and improve energy resolution in challenging experimental contexts, providing a solid foundation for future refinements and broader applications.