Speaker
Description
Accurate neutron/gamma pulse shape discrimination (PSD) in plastic scintillators is strongly limited at low energies and further complicated by the scarcity and uncertainty of labeled training data, particularly for mixed neutron–gamma sources such as Cf-252. Conventional supervised deep learning approaches rely heavily on clean labels, which are often difficult to obtain in practical experimental conditions.
In this work, we propose a waveform-based neutron/gamma discrimination framework based on self-supervised and semi-supervised learning to reduce label dependence and improve robustness. A self-supervised pretraining stage is first employed to learn latent representations from large volumes of unlabeled scintillation waveforms using contrastive learning. The pretrained model is then fine-tuned with a limited set of labeled neutron and gamma events acquired from Cf-252 and Co-60 sources. Experimental results using an EJ-276 plastic scintillation detector show that the proposed approach significantly improves discrimination performance in the low-energy region below 150 keVee, achieving higher Figures of Merit compared to fully supervised models trained on the same labeled dataset
| Minioral | Yes |
|---|---|
| IEEE Member | Yes |
| Are you a student? | No |