[Paper Review] Machine Learning for Track Finding at PANDA

Waleed Esmail, T. Stockmanns+1 more|arXiv (Cornell University)|Oct 16, 2019

Image Processing and 3D Reconstruction5 references4 citations

TL;DR

This paper proposes a deep learning-based track finding algorithm for the PANDA Forward Tracking System (FTS), using a three-stage approach: a feedforward neural network to create 2D track segments, geometric 3D extension using skewed layers, and an LSTM-RNN to connect segments into full tracks. The method achieves >95% purity and >90% efficiency for moderate to high momentum tracks (3.5–6.5 GeV/c), with 80% efficiency for low-momentum tracks (1 GeV/c).

ABSTRACT

We apply deep learning methods as a track finding algorithm to the PANDA Forward Tracking Stations (FTS). The problem is divided into three steps: The first step relies on an Artificial Neural Network (ANN) that is trained as a binary classifier to build track segments in three different parts of the FTS, namely FT1,FT2, FT3,FT4, and FT5,FT6. The ANN accepts hit pairs as an input and outputs a probability that they are on the same track or not. The second step builds 3D track segments from the 2D ones and is based on the geometry of the detector. The last step is to match the track segments from the different parts of the FTS to form a full track candidate, and is based on a Recurrent Neural Network (RNN). The RNN is used also as a binary classifier that outputs the probability that the combined track segments are a true track or not. The performance of the algorithm is judged based on the purity, efficiency and the ghost ratio of the reconstructed tracks. The purity specifies which fraction of hits in one track come from the correct particle. The correct particle is the particle, which produces the majority of hits in the track. The efficiency is defined as the ratio of the number of correctly reconstructed tracks to all generated tracks.

Motivation & Objective

• Develop a pattern recognition algorithm for reconstructing charged particle tracks in the PANDA Forward Tracking System (FTS) using deep learning.
• Address the challenge of track finding in a complex, multi-layered detector with varying spatial resolution and geometry.
• Improve track reconstruction efficiency and purity for low, moderate, and high momentum particles in the FTS.
• Demonstrate the feasibility of using deep learning—particularly RNNs—for connecting segmented tracks into full particle trajectories.
• Optimize the algorithm for real-time performance and robustness in high-rate, high-multiplicity environments typical of PANDA.

Proposed method

• Train a feedforward Artificial Neural Network (ANN) as a binary classifier to identify hit pairs belonging to the same track in three FTS regions: (FT1+FT2), (FT3+FT4), and (FT5+FT6).
• Use geometric constraints from the detector's skewed layer configuration to extend 2D track segments into 3D by projecting hits using drift time and spatial coordinates.
• Apply a Long Short-Term Memory (LSTM) recurrent neural network to sequence the 3D track segments across stations, modeling the temporal order of hits along a track.
• Train the LSTM to output a probability score indicating whether a sequence of track segments constitutes a true particle track, using corrected hit positions (X, Y, Z) as input.
• Implement dropout (50%) and stacking of three LSTM layers to prevent overfitting and improve generalization during training.
• Use Monte Carlo simulations from PandaRoot to generate training data with isotropic muon tracks (1–10 GeV/c) and corresponding detector hits (wire position and isochrone radius).

Experimental results

Research questions

• RQ1 Can deep learning effectively reconstruct charged particle tracks in the PANDA FTS with high purity and efficiency?
• RQ2 How well can a feedforward ANN classify hit pairs as belonging to the same track in segmented detector regions?
• RQ3 To what extent does geometric modeling of skewed layers improve 3D track reconstruction from 2D segments?
• RQ4 Can an LSTM-RNN effectively connect track segments across multiple FTS stations into a coherent full-track candidate?
• RQ5 What is the performance trade-off between track purity, efficiency, and ghost ratio across different momentum regimes?

Key findings

• For high-momentum tracks (6.5 GeV/c), the algorithm achieves nearly 100% purity and efficiency, indicating excellent reconstruction quality.
• At moderate momentum (3.5 GeV/c), purity exceeds 95% and efficiency remains above 90%, demonstrating strong performance under realistic conditions.
• For low-momentum tracks (1 GeV/c), efficiency drops to 80% for tracks with 5 hits, indicating challenges in low-multiplicity, low-momentum regimes.
• The LSTM-RNN model reaches 95% validation accuracy after 1000 training iterations, confirming its ability to learn complex sequential dependencies in track segments.
• The 3D reconstruction using skewed layers improves spatial resolution, with vertical resolution (~0.15 cm) matching the design goal and horizontal resolution being 10× better.
• Ghost ratio is minimized through the use of geometric constraints and the LSTM’s sequence modeling, reducing false track candidates.