Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Particle Identification with MLPs and PINNs Using HADES Data

Marvin Kohls

TL;DR

This work tackles the challenge of obtaining high-purity particle-identification samples in HADES data by integrating physics constraints into neural networks. It introduces a physics-informed neural network framework with domain-adversarial training that maps simulated and experimental data into a shared latent space while enforcing a Bethe-Bloch-based physics constraint in the loss: $L = L_class(y,y_true) + \alpha L_domain(d_sim,0) + \alpha L_domain(d_exp,1) + \lambda L_Bethe-Bloch(dE/dx, dE/dx_theory)$. A Kaon-focused PID study demonstrates that incorporating physics information into the domain-adversarial training substantially improves PID performance, achieving AUC values around 0.82 for DANN and 0.85 for MLP setups. The approach reduces simulation bias and promises more reliable analyses for hadron identification at few-GeV energies, offering a path toward more accurate interpretation of rare-event samples in nuclear and particle physics.

Abstract

In experimental nuclear and particle physics, the extraction of high-purity samples of rare events critically depends on the efficiency and accuracy of particle identification (PID). In this work, we present a PID method applied to HADES data at the level of fully reconstructed particle track candidates. The results demonstrate a significant improvement in PID performance compared to conventional techniques, highlighting the potential of physics-informed neural networks as a powerful tool for future data analyses.

Particle Identification with MLPs and PINNs Using HADES Data

TL;DR

This work tackles the challenge of obtaining high-purity particle-identification samples in HADES data by integrating physics constraints into neural networks. It introduces a physics-informed neural network framework with domain-adversarial training that maps simulated and experimental data into a shared latent space while enforcing a Bethe-Bloch-based physics constraint in the loss: . A Kaon-focused PID study demonstrates that incorporating physics information into the domain-adversarial training substantially improves PID performance, achieving AUC values around 0.82 for DANN and 0.85 for MLP setups. The approach reduces simulation bias and promises more reliable analyses for hadron identification at few-GeV energies, offering a path toward more accurate interpretation of rare-event samples in nuclear and particle physics.

Abstract

In experimental nuclear and particle physics, the extraction of high-purity samples of rare events critically depends on the efficiency and accuracy of particle identification (PID). In this work, we present a PID method applied to HADES data at the level of fully reconstructed particle track candidates. The results demonstrate a significant improvement in PID performance compared to conventional techniques, highlighting the potential of physics-informed neural networks as a powerful tool for future data analyses.

Paper Structure

This paper contains 6 sections, 1 equation, 3 figures.

Table of Contents

  1. The HADES Experiment
  2. Particle Identification
  3. Short-Lived Particles
  4. Charged Hadron Identification with Neural Networks
  5. Results
  6. Discussion and Conclusion

Figures (3)

  • Figure 1: Exploded view of he HADES experimental setup.
  • Figure 2: Setup of the physics informed neural network for particle identification.
  • Figure 3: Performance for different ANN setups in identifying $K^{+}$ in Ag(1.58 A GeV)+Ag data. PI stands for "Physics Informed" neural networks.