Speaker
Description
High-speed digital Data Acquisition (DAQ) systems in modern nuclear physics face a common bottleneck: the massive data throughput generated by digitizing full waveforms at high sampling rates. This limitation restricts the effective counting rate and increases the beam time required to achieve statistical significance.
This work proposes a generic, deep learning-based pulse compression framework designed for scintillation detectors. The approach utilizes a 1D Convolutional Autoencoder (1D-CAE) with an asymmetrical architecture optimized for Edge Computing. While this study presents the validation of the model architecture, the ultimate objective is the hardware implementation of the Encoder directly into the front-end electronics (e.g., FPGAs) for real-time operation.
To validate this concept, the method was tested using data from the Neutron Detector Array (NEDA). Results demonstrate that the proposed architecture successfully compresses pulses into a lower-dimensional latent space while preserving the critical morphological features required for Pulse Shape Discrimination (PSD). This validation provides the necessary proof-of-concept to proceed with the firmware deployment, aiming to significantly reduce data bandwidth without compromising the intrinsic detection efficiency.
| Minioral | Yes |
|---|---|
| IEEE Member | No |
| Are you a student? | No |