Particle Identification with Deep Neural Networks Across Collision Energies in Simulated Proton-Proton Collisions

TL;DR

This paper demonstrates that a deep neural network trained on simulated high-energy proton-proton collision data can accurately identify particle types across different collision energies without retraining, indicating the model captures physically meaningful features.

Contribution

It introduces a proof-of-concept deep learning approach for particle identification that generalizes across collision energies without transfer learning or fine-tuning.

Findings

Model maintains >91% accuracy across different collision energies.

High accuracy (>96%) achieved for high transverse momentum particles.

Model captures physically meaningful features beyond simple kinematic patterns.

Abstract

This study demonstrates a proof-of-concept application of a deep neural network for particle identification in simulated high transverse momentum proton-proton collisions, with a focus on evaluating model performance under controlled conditions. A model trained on simulated Large Hadron Collider (LHC) proton-proton collisions at $\sqrt{s} = 13\,\mathrm{TeV}$ is used to classify nine particle species based on seven kinematic-level features. The model is then tested on simulated high transverse momentum Relativistic Heavy Ion Collider (RHIC) data at $\sqrt{s} = 200\,\mathrm{GeV}$ without any transfer learning, fine-tuning, or weight adjustment. It maintains accuracy above 91% for both LHC and RHIC sets, while achieving above 96% accuracy for all RHIC sets, including the $p_T > 7\,\mathrm{GeV}/c$ set, despite never being trained on any RHIC data. Analysis of per-class accuracy reveals how quantum chromodynamics (QCD) effects, such as leading particle effect and kinematic overlap at high $p_T$, shape the model's performance across particle types. These results suggest that the model captures physically meaningful features of high-energy collisions, rather than simply overfitting to kinematics of the training data. This study demonstrates the potential of simulation-trained deep neural networks to remain effective across lower energy regimes within a controlled environment, and motivates further investigation in realistic settings using detector-level features and more advanced network architectures.