Spectroscopic Pulse Embeddings by Contrastive Learning from Unlabeled Data for Pile-Up Analysis.

Area of Science:

• Nuclear physics
• Signal processing
• Machine learning

Background:

• Pile-up effect in nuclear spectroscopy distorts measurements due to overlapping detector pulses.
• Current deep learning methods struggle with real-world data due to simulation-experiment discrepancies.

Purpose of the Study:

• To develop a robust and transferable deep learning framework for pile-up correction using unlabeled real nuclear pulse signals.
• To improve the accuracy of nuclear measurements, especially under high counting-rate conditions.

Main Methods:

• Utilized a contrastive learning framework on unlabeled real nuclear pulse signals.
• Employed a zero-crossing-based strategy for pulse segmentation and physics-inspired data augmentations.
• Used a 1D ResNet encoder for representation learning and transferred these to specific tasks.

Main Results:

• Demonstrated strong performance and robustness in real nuclear radiation detection systems.
• Achieved significant improvements in pile-up identification and counting-rate estimation.
• Showcased particular advantages in challenging peak pile-up scenarios.

Conclusions:

• The proposed contrastive learning method offers a robust solution for pile-up correction in nuclear spectroscopy.
• This approach overcomes limitations of supervised methods relying on simulated data.
• The framework provides transferable representations for enhanced nuclear measurement accuracy.