A Deep Neural Network for Pixel-Level Electromagnetic Particle Identification in the MicroBooNE Liquid Argon Time Projection Chamber (1808.07269v1)

Abstract: We have developed a convolutional neural network (CNN) that can make a pixel-level prediction of objects in image data recorded by a liquid argon time projection chamber (LArTPC) for the first time. We describe the network design, training techniques, and software tools developed to train this network. The goal of this work is to develop a complete deep neural network based data reconstruction chain for the MicroBooNE detector. We show the first demonstration of a network's validity on real LArTPC data using MicroBooNE collection plane images. The demonstration is performed for stopping muon and a $\nu_\mu$ charged current neutral pion data samples.

• The paper demonstrates a novel U-ResNet CNN that segments electromagnetic particles at the pixel level, achieving misclassification rates as low as 3.9% on simulated νμ interactions.
• It employs transfer learning and class/pixel-wise loss weighting to counter class imbalance in LArTPC images, thereby stabilizing and expediting the training process.
• Validation on real detector data, including Michel electron and charged-current π0 events, confirms the network’s efficacy for accurate and automated neutrino event reconstruction.

A Deep Neural Network for Pixel-Level Electromagnetic Particle Identification in the MicroBooNE Liquid Argon Time Projection Chamber

This paper presents the design, implementation, and analysis of a convolutional neural network (CNN) for achieving pixel-level discrimination of electromagnetic (EM) particles within images captured by the MicroBooNE Liquid Argon Time Projection Chamber (LArTPC) detector. The MicroBooNE detector operates using liquid argon to capture high-resolution particle interaction images, instrumental for physics analyses in neutrino research. The advancements described in the paper mark critical developments towards automating and enhancing particle identification, bridging the gap between raw detector data and physics analysis.

Technical Contributions and Methodology

The authors elaborate on the development and application of a U-ResNet architecture, a synthesis of U-Net and Residual Network (ResNet) designs. U-ResNet is tailored for image segmentation tasks and optimizes pixel-level classification by leveraging both hierarchical feature representations and spatial information retention through its unique encoding and decoding paths, combined with residual learning modules.

Key technical enhancements in the CNN training include:

1. Transfer Learning: Initially training the network on an image classification task to expedite and stabilize the subsequent labeling task.
2. Class/Pixel-wise Loss Weighting: Introducing both class-wise and pixel-wise loss weighting mechanisms to address the significant class imbalance inherent in LArTPC images, where background pixels predominantly outnumber signal pixels.

The training utilized simulated particle interactions generated by the MultiPartVertex (MPV) algorithm, allowing for controlled variations in particle types and topologies. This simulated dataset serves the dual purpose of training and validating the CNN's performance before deployment on real detector data.

Performance Evaluation

The U-ResNet’s performance was quantified using several metrics on both simulated and real-world data. Noteworthy benchmarks include:

• Incorrectly Classified Pixel Fraction (ICPF): The network achieved an ICPF of 6.0% and 3.9% on $\nu_e$ and $\nu_\mu$ simulated interactions, respectively.
• Shower and Track Error Rates: Robust performance with track error rates as low as 2.3% in low-energy scenarios, demonstrating the network's efficacy across a spectrum of particle energies and configurations.

Validation on Detector Data

A robust validation was conducted on actual detector data, comparing neural network outcomes against physicist-curated labels. Two specific datasets were scrutinized:

1. Michel Electron Events: Known for clear track and shower delineation, these events enabled precise benchmarking, showcasing the network's ability to maintain low disagreement rates around 1.9%.
2. Charged-Current $\pi^0$ Events: Featuring more complex topologies, the network was able to classify particle interactions effectively, with an observed disagreement fraction of 3.4%.

The validation evidence underscored the network's applicability and reliability, extending beyond synthetic data to real-world observations.

Implications and Future Directions

Theoretical Implications: The successful implementation of U-ResNet underscores the value of deep learning in neutrino physics, notably in LArTPC-based experiments. The ability to automate and accurately classify particle interactions on a pixel-level augments the overall precision of neutrino event reconstructions.

Practical Implications: From a practical standpoint, the integration of such networks can streamline data analysis pipelines, reducing the bottle-neck posed by manual labeling and opening avenues for real-time data processing in high-rate environments. The potential scalability to other detectors, such as those envisioned for the Deep Underground Neutrino Experiment (DUNE), marks a significant stride.

Future Work: Future developments may include enhancing the network to harness multi-plane projection data for improved spatial resolution and trajectory reconstruction. The continued calibration refinement, harmonizing the data and simulation pixel scales, holds promise for further lowering systematic discrepancies.

In conclusion, this paper delineates a comprehensive approach to employing deep learning for particle identification in LArTPCs, setting a foundational benchmark for subsequent innovations in neutrino detection and analysis. The methodologies and results presented affirm the capability of deep neural networks to transform particle physics data interpretation, enhancing both theoretical research and practical experimentations.