Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Using deep neural networks to improve the precision of fast-sampled particle timing detectors

Mateusz Kocot, Krzysztof Misan, Valentina Avati, Edoardo Bossini, Leszek Grzanka, Nicola Minafra

TL;DR

Time walk from charge fluctuations degrades timing precision in fast particle timing detectors. The authors compare multilayer perceptrons, convolutional networks, and UNet models trained on DESY-II test-beam data with MCP-PMT ground truth, showing UNet delivers the best timing precision and consistently outperforms the CFD baseline by about 8%–23%. They demonstrate robustness to reduced-sample LHC-like data (24 samples) with timing precision around 62 ps, and observe similar gains across multiple detector channels, highlighting practical impact for CMS-PPS timing improvements. The approach supports online processing with low inference latency and provides a data-driven path to replace or augment CFD in future high-energy physics timing systems.

Abstract

Measurements from particle timing detectors are often affected by the time walk effect caused by statistical fluctuations in the charge deposited by passing particles. The constant fraction discriminator (CFD) algorithm is frequently used to mitigate this effect both in test setups and in running experiments, such as the CMS-PPS system at the CERN's LHC. The CFD is simple and effective but does not leverage all voltage samples in a time series. Its performance could be enhanced with deep neural networks, which are commonly used for time series analysis, including computing the particle arrival time. We evaluated various neural network architectures using data acquired at the test beam facility in the DESY-II synchrotron, where a precise MCP (MicroChannel Plate) detector was installed in addition to PPS diamond timing detectors. MCP measurements were used as a reference to train the networks and compare the results with the standard CFD method. Ultimately, we improved the timing precision by 8% to 23%, depending on the detector's readout channel. The best results were obtained using a UNet-based model, which outperformed classical convolutional networks and the multilayer perceptron.

Using deep neural networks to improve the precision of fast-sampled particle timing detectors

TL;DR

Time walk from charge fluctuations degrades timing precision in fast particle timing detectors. The authors compare multilayer perceptrons, convolutional networks, and UNet models trained on DESY-II test-beam data with MCP-PMT ground truth, showing UNet delivers the best timing precision and consistently outperforms the CFD baseline by about 8%–23%. They demonstrate robustness to reduced-sample LHC-like data (24 samples) with timing precision around 62 ps, and observe similar gains across multiple detector channels, highlighting practical impact for CMS-PPS timing improvements. The approach supports online processing with low inference latency and provides a data-driven path to replace or augment CFD in future high-energy physics timing systems.

Abstract

Measurements from particle timing detectors are often affected by the time walk effect caused by statistical fluctuations in the charge deposited by passing particles. The constant fraction discriminator (CFD) algorithm is frequently used to mitigate this effect both in test setups and in running experiments, such as the CMS-PPS system at the CERN's LHC. The CFD is simple and effective but does not leverage all voltage samples in a time series. Its performance could be enhanced with deep neural networks, which are commonly used for time series analysis, including computing the particle arrival time. We evaluated various neural network architectures using data acquired at the test beam facility in the DESY-II synchrotron, where a precise MCP (MicroChannel Plate) detector was installed in addition to PPS diamond timing detectors. MCP measurements were used as a reference to train the networks and compare the results with the standard CFD method. Ultimately, we improved the timing precision by 8% to 23%, depending on the detector's readout channel. The best results were obtained using a UNet-based model, which outperformed classical convolutional networks and the multilayer perceptron.
Paper Structure (21 sections, 12 figures, 2 tables)

This paper contains 21 sections, 12 figures, 2 tables.

Table of Contents

  1. Introduction
  2. State of the art
  3. Analytical approaches
  4. Machine learning methods
  5. Dataset
  6. Data source
  7. Preprocessing
  8. Final dataset
  9. Constant Fraction Discriminator
  10. Network architectures
  11. Training configuration
  12. Hyperparameter tuning procedure
  13. Multilayer perceptron
  14. Convolutional Network
  15. UNet
  16. ...and 6 more sections

Figures (12)

  • Figure 1: Data flow diagram for the analysis of the timing data (blue: electronics, green: data, red: digital algorithm). Multiple algorithms can be used to retrieve the time of arrival from the waveform data. In this research we focus on digital algorithms working in the offline mode.
  • Figure 2: Time walk error illustrated as a difference in the threshold crossing times between two signals with the same shapes but different amplitudes
  • Figure 3: Example waveforms from the MCP (left) and a diamond detector (right).
  • Figure 4: Examples of noisy (left) and saturated (right) events acquired using the diamond sensor
  • Figure 5: Maximum voltage histograms for MCP (left) and two selected diamond detectors (middle and right). The red lines on the histograms of the diamond detectors indicate the minimum and maximum amplitude cuts. It is important to note that the histograms look different for each detector, and therefore, it was necessary to find the amplitude cut values separately for each of them.
  • ...and 7 more figures