Deep Learning Based Pile-Up Correction Algorithm for Spectrometric Data Under High-Count-Rate Measurements.

Area of Science:

• Nuclear Science and Engineering
• Applied Physics
• Data Science

Background:

• Gamma-ray spectroscopy is crucial for identifying radioactive materials via energy spectrum analysis.
• High count rates in spectroscopy cause pile-up effects, distorting spectra and hindering accurate analysis.
• Automated and precise pile-up correction methods are needed for reliable nuclear material characterization.

Purpose of the Study:

• To develop a novel deep learning (DL) framework for accurate energy spectrum recovery in gamma-ray spectroscopy.
• To address and mitigate spectral distortions caused by pile-up effects under high count rate conditions.
• To enhance the accuracy of isotope identification and activity estimation in nuclear science applications.

Main Methods:

• A 2D attention U-Net deep learning model was employed for energy spectrum recovery.
• Count rate information of pile-up signals was integrated into the DL framework.
• An Energy-Duration matrix from preprocessed pulse signals served as model input, extracting temporal and spatial features.
• Training data were generated using an open-source simulator and a public gamma spectrum database.

Main Results:

• The proposed DL framework effectively predicted accurate energy spectra, minimizing errors.
• Performance was validated using Kullback-Leibler divergence, Mean Squared Error, Energy Resolution, and Full Width at Half Maximum.
• The model demonstrated robustness and accuracy even under severe pile-up effects and high count rates.

Conclusions:

• The developed framework offers a robust solution for automated pile-up correction in gamma-ray spectroscopy.
• This approach significantly improves the accuracy of spectrum estimation for high-activity nuclear analysis.
• The integration of temporal and spatial learning shows promise for advancing nuclear measurement techniques.