Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

PE-GAN: Prior Embedding GAN for PXD images at Belle II

Baran Hashemi, Nikolai Hartmann, Thomas Kuhr, Martin Ritter, Matej srebre

TL;DR

Belle II's Pixel Vertex Detector faces significant background contamination that, if stored as raw overlay data, demands prohibitive storage and bandwidth. The authors present an on-demand background generator based on a class-conditioned GAN with sensor-ID conditioning and contrastive learning to produce 40 sensor-dependent PXD hitmaps per event. The model evolves from BigGAN-deep to a SAGAN-based, contrastive conditioned architecture with consistency regularization to achieve high fidelity and training stability, generating hitmaps in about 6 seconds per event on CPU. Validation against Geant4 shows close agreement in pixel intensities, occupancies, and impact-parameter resolutions, supporting the use of generated backgrounds in Belle II simulations and saving storage resources.

Abstract

The pixel vertex detector (PXD) is an essential part of the Belle II detector recording particle positions. Data from the PXD and other sensors allow us to reconstruct particle tracks and decay vertices. The effect of background hits on track reconstruction is simulated by adding measured or simulated background hit patterns to the hits produced by simulated signal particles. This model requires a large set of statistically independent PXD background noise samples to avoid a systematic bias of reconstructed tracks. However, data from the fine-grained PXD requires a substantial amount of storage. As an efficient way of producing background noise, we explore the idea of an on-demand PXD background generator using conditional Generative Adversarial Networks (GANs) with contrastive learning, adapted by the number of PXD sensors in order to both increase the image fidelity and produce sensor-dependent PXD hitmaps.

PE-GAN: Prior Embedding GAN for PXD images at Belle II

TL;DR

Belle II's Pixel Vertex Detector faces significant background contamination that, if stored as raw overlay data, demands prohibitive storage and bandwidth. The authors present an on-demand background generator based on a class-conditioned GAN with sensor-ID conditioning and contrastive learning to produce 40 sensor-dependent PXD hitmaps per event. The model evolves from BigGAN-deep to a SAGAN-based, contrastive conditioned architecture with consistency regularization to achieve high fidelity and training stability, generating hitmaps in about 6 seconds per event on CPU. Validation against Geant4 shows close agreement in pixel intensities, occupancies, and impact-parameter resolutions, supporting the use of generated backgrounds in Belle II simulations and saving storage resources.

Abstract

The pixel vertex detector (PXD) is an essential part of the Belle II detector recording particle positions. Data from the PXD and other sensors allow us to reconstruct particle tracks and decay vertices. The effect of background hits on track reconstruction is simulated by adding measured or simulated background hit patterns to the hits produced by simulated signal particles. This model requires a large set of statistically independent PXD background noise samples to avoid a systematic bias of reconstructed tracks. However, data from the fine-grained PXD requires a substantial amount of storage. As an efficient way of producing background noise, we explore the idea of an on-demand PXD background generator using conditional Generative Adversarial Networks (GANs) with contrastive learning, adapted by the number of PXD sensors in order to both increase the image fidelity and produce sensor-dependent PXD hitmaps.
Paper Structure (9 sections, 5 equations, 5 figures, 2 tables)

This paper contains 9 sections, 5 equations, 5 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Pixel Vertex Detector
  3. The Problem
  4. PXD Background Generation
  5. The Model
  6. Model Validation
  7. Training Evaluation Methods
  8. Validation Methods
  9. Summary and outlook

Figures (5)

  • Figure 1: Comparison of color-reversed PXD image (up) and generated image (down)
  • Figure 2: First singular values of the discriminator for each layer in (a) a model with collapse (b) our model. The most important thing to notice here is the sudden drop in the values over a absolute large range which is an indication of discriminator's overfitting.
  • Figure 3: Pixel intensity value comparison. Obviously a large portion of the images are totally dark. There is a very narrow discrepancy between the Geant4 simulated and GAN generated pixel values.
  • Figure 4: Sensor dependent mean occupancy comparison between Geant4 simulated and GAN generated images
  • Figure 5: Comparison of the resolution of the signed distance in the transverse plane from the pivotal point to the helix, between simulated and generated background and no background