TL;DR
This paper introduces a convolutional neural network (CNN) for classifying particle energy deposits in the ProtoDUNE-SP liquid argon time projection chamber. The CNN effectively distinguishes between track-like (e.g., muons, protons) and shower-like (e.g., electrons, photons) energy deposits, and also identifies Michel electrons, crucial for precise neutrino interaction reconstruction in the future DUNE experiment. The algorithm demonstrates high efficiency and excellent agreement between experimental data and simulations.
What Problem Does This Paper Address?
Reconstructing neutrino interactions in liquid argon time projection chambers (LArTPCs) like ProtoDUNE-SP requires accurately identifying and classifying the resulting particles. Distinguishing between "track-like" particles (which travel in relatively straight lines) and "shower-like" particles (which spread out due to electromagnetic interactions) is a fundamental challenge. This distinction is vital for determining particle types, their energies, and ultimately, the nature of the neutrino interaction. Furthermore, identifying Michel electrons, decay products of muons, provides additional information for particle identification and reconstruction. Existing methods often struggle with this fine-grained classification, especially in the complex environment of a LArTPC.
Methods
The authors propose a convolutional neural network (CNN) approach for classifying individual energy deposits (hits) within the ProtoDUNE-SP detector. Instead of analyzing the entire detector image, the CNN focuses on small, localized "patches" of data (48x48 pixels) centered around each hit. This patch-based approach is designed for computational efficiency, allowing the algorithm to run on standard CPUs without specialized GPUs during inference.
The CNN architecture is relatively simple, featuring a single convolutional layer to extract features, followed by two dense layers. The output is split into two branches: one for classifying hits as track, shower, or empty (TSE), and another for identifying Michel electrons. The TSE branch uses a softmax activation function, while the Michel branch uses a sigmoid function. The training process involved a large dataset of simulated events, and the performance was evaluated using separate training, validation, and test sets.
Datasets
The CNN was trained and evaluated using data from the ProtoDUNE-SP detector. This includes:
- Simulated Data: Used for training the CNN. Approximately 30 million images were generated, with specific samples for track, shower, empty, and Michel electron classes. Detector effects like space charge were included in the simulation.
- Experimental Data: Used for validation and performance evaluation. This includes data from:
- Cosmic ray muons.
- Charged particle beam data (pions, muons, protons, positrons) at 1 GeV/c from the H4-VLE test beam at CERN.
Key Insights and Ideas
- Patch-based CNN for Efficiency: The decision to use small, localized image patches for the CNN is a key insight. This significantly reduces computational load, making the algorithm suitable for real-time analysis on standard hardware, which is crucial for large-scale experiments like DUNE.
- Dual-Branch Output: Separating the TSE classification from the Michel electron classification allows for tailored optimization and addresses the different characteristics and challenges of each task.
- Good Agreement Between Data and Simulation: A critical finding is the strong agreement between the CNN's performance on simulated data and its performance on real experimental data. This indicates that the simulation accurately captures the detector physics and that the trained algorithm is robust and generalizable.
- Michel Electron Identification is Challenging but Important: The paper highlights that identifying Michel electrons is more difficult due to their varied appearance and potential confusion with other phenomena. However, their accurate identification is crucial for disentangling muon and pion interactions.
Key Results
- High Efficiency: The CNN achieves high efficiency in classifying hits as track-like, shower-like, or Michel-like.
- Strong Separation: The shower classifier shows a clear separation between shower hits and other hits, with an F1 score optimized at 0.72 for hit-level classification.
- Particle-Level Performance: For particle-level classification, a threshold of 0.81 on the average shower score leads to excellent agreement between data and simulation, with fractional differences within 1-2%.
- Michel Electron Classification: While challenging, the Michel electron classifier shows promise, with a hit-level classification threshold of 0.19 yielding good results. The particle-level classification for muons and pions also shows good agreement between data and simulation.
Limitations and Future Work
- Michel Electron Classification Challenges: The paper acknowledges that Michel electron classification is more challenging, partly due to the large variation in Michel electron interactions and potential confusion with other particle signatures. This leads to a less optimal performance in terms of F1 metric compared to shower classification.
- Future Work: The authors suggest that combining this CNN with clustering algorithms can help select high-purity samples of Michel electron events. The algorithm is already being integrated into ongoing ProtoDUNE-SP analyses for cross-section measurements and detector calibration. The robustness of the algorithm across different particle species and its consistency between data and simulation suggest its broad applicability in future LArTPC analyses.
Summary updated