Source-linked AI summary
A Convolutional Neural Network Neutrino Event Classifier
A. Aurisano, A. Radovic, D. Rocco, A. Himmel, M. D. Messier, E. Niner, G. Pawloski, F. Psihas, A. Sousa, P. Vahle
TL;DR
The paper addresses signal-background categorization in high-energy particle detectors, where reconstruction errors and predefined features limit classification. It develops and evaluates the CNN-based CVN algorithm on simulated NOvA events, achieving strong separation with minimal reconstruction and improved νe CC efficiency over a cited algorithm.
Problem
HEP event classification commonly depends on reconstructed features, which can introduce errors and restrict inputs to features previously designed for the experiment.
Method
The paper applies convolutional neural networks to detector energy-deposition images and develops CVN for simulated NOvA neutrino events.
Results
CVN achieved 49% νe CC signal-detection-optimized efficiency versus 35% for the cited baseline, while achieving excellent signal-background separation with minimal reconstruction.
Takeaways & Limitations
CVN demonstrates that CNNs can work effectively on non-natural sampling-calorimeter images and may transfer across detector technologies and analyses.
Takeaways & Limitations
The reported misidentification problem remains relevant because NC interactions can resemble CC interactions when pion activity imitates leptonic activity.
Abstract
from arXiv · showhide
Convolutional neural networks (CNNs) have been widely applied in the computer vision community to solve complex problems in image recognition and analysis. We describe an application of the CNN technology to the problem of identifying particle interactions in sampling calorimeters used commonly in high energy physics and high energy neutrino physics in particular. Following a discussion of the core concepts of CNNs and recent innovations in CNN architectures related to the field of deep learning, we outline a specific application to the NOvA neutrino detector. This algorithm, CVN (Convolutional Visual Network) identifies neutrino interactions based on their topology without the need for detailed reconstruction and outperforms algorithms currently in use by the NOvA collaboration.
1 Introduction
The paper frames particle-interaction categorization as separating signal from background while avoiding errors and feature limitations introduced by conventional reconstruction. It applies CNNs to detector readouts that function as images and develops CVN for simulated NOvA events.
- Motivation: Conventional HEP classification reconstructs clusters, tracks, showers, jets, or rings before applying selections or machine-learning algorithms.These reconstructed objects are summarized by energies, directions, and shapes.
- Motivation: Reconstruction mistakes can cause incorrect event categorization, while engineered inputs exclude features not previously imagined and implemented.
- CNN motivation: CNNs address an analogous computer-vision problem by learning features rather than relying on specifically constructed features.
- Detector representation: Sampling calorimeters and other detectors record energy deposits across small regions, producing interaction images suited to computer-vision analysis.Prior studies with Daya Bay data and simulated LHC jets demonstrated CNN power in high-energy physics.
- Paper scope: The paper describes CNN construction and training, then tests CVN on simulated NOvA events and benchmarks its performance.CVN is the paper’s Convolutional Visual Network application.
2 Deep Learning and Convolutional Neural Networks
CNNs learn hierarchical image features through convolutions, nonlinearities, pooling, and regularization, addressing limitations of traditional multilayer perceptrons. The paper also motivates a GoogLeNet-inspired architecture for detector-image classification.
- Traditional neural networks: Traditional MLPs use fully connected layers to approximate mappings from n-dimensional inputs to m-dimensional outputs using supervised learning.Weights and biases are learned from examples paired with known ground-truth outputs.
- MLP limitations: MLPs scale poorly to raw inputs because they depend on difficult, time-consuming feature extraction, while deep sigmoid networks can be difficult to train.Sigmoid saturation can drive gradients toward zero and slow or halt learning.
- Deep-learning innovations: CNNs automatically learn robust features, while ReLU nonlinearities and dropout support efficient training and reduce over-training.ReLU is defined as f(x) = max(0, x); dropout randomly removes connections during training.
- CNN representation: A CNN represents detector data with height, width, and channel dimensions, analogous to image channels.
- Convolution: Convolution applies learned four-dimensional filters across input channels to produce feature maps containing alternative learned representations.Early feature maps can resemble the input, whereas later maps become more abstract.
- Pooling: Pooling downsamples feature maps using regional maxima or averages, reducing computation and sensitivity to absolute feature location.Overlapping pooled regions can reduce information loss.
- Architecture: The described network is inspired by GoogLeNet and uses network-in-network structures to increase convolutional learning capacity while reducing dimensionality.
3 Application to NOvA event classification
CVN applies a two-view convolutional neural network to simulated NOvA detector events, classifying neutrino interactions from detector topology across four flavor-interaction categories. The architecture processes x−z and y−z views separately, then combines them for classification.
- CVN architecture: Inception modules combine 1 × 1, 3×3, and 5×5 convolutions with overlapping pooling, concatenating branch outputs into the next layer.The module maintains the input feature-map dimensions while applying filters at different scales.
- Event categories: CVN characterizes candidate neutrino events as νµ CC, νe CC, ντ CC, or ν NC interactions.Charged-current categories contain a charged lepton and hadronic activity, whereas only hadronic activity is visible for neutral-current interactions.
- Event representation: The input is formed from clustered energy deposits and two detector-view grids covering 100 planes by 80 cells.Separate grids represent the x−z and y−z views after deposits are clustered in space and time into slices.
- Event representation: 8-bit encoding saturates hits above 278 MeV while reducing storage requirements by a factor of eight without significantly compromising representational capacity.The encoded detector activity serves as an image-like input to the neural network.
- CVN architecture: The network sends the X and Y views through parallel GoogLeNet-based architectures before merging them through a final inception module and softmax output.Each branch uses successive convolution, pooling, and local response normalization, with inception modules in downstream layers.
- Training: Training used 4.7 million simulated NOvA Far Detector events, split 80% for training and 20% for testing, over a full week of GPU hours.The similar training and test losses throughout training indicated that the network had not overtrained.
4 Results
CVN was evaluated on an independent simulated NOvA sample using classifier outputs to select νe CC and νµ CC interactions. It achieved strong signal-background separation, with νe CC efficiency substantially exceeding the existing benchmark while systematic sensitivity remained comparable.
- Evaluation setup: An independent simulated NOvA interaction sample was used to assess CVN performance after training and testing.The sample was weighted by the simulated beam flux and oscillation probabilities.
- Classifier performance: CVN classifier outputs were converted into νe CC and νµ CC classifications by summing the relevant softmax component probabilities.Cumulative efficiency, purity, and their product were evaluated as functions of the classification threshold.
- Classifier performance: CVN achieved excellent separation of signal and background, with the main remaining νe CC background coming from intrinsic beam electron neutrinos.CVN does not distinguish oscillation-produced νe CC events from promptly produced beam νe CC events because they differ only in energy distributions.
- Optimization: CVN provided selection criteria optimized either for signal purity or measurement efficiency for both νe CC and νµ CC events.The two Figure of Merit choices prioritize pure signal samples or efficient samples for parameter measurements, respectively.
- Comparison with existing algorithms: 49% νe CC signal-detection-optimized efficiency exceeded the 35% efficiency of the existing algorithm, while 58% νµ CC measurement-optimized efficiency compared with 57%.CVN purity closely matched the purities of the previously reported samples.
- Systematic uncertainties: CVN showed no increased sensitivity to the studied systematic uncertainties compared with the selectors used in earlier NOvA analyses.Calibration and scintillator saturation variations produced small, comparable shifts in signal and background selections.
5 Conclusion and future work
The paper concludes that CVN can classify neutrino events effectively with minimal reconstruction, using calorimeter readouts as non-natural images. Future work extends the approach to additional interaction types, particle identification, and semantic segmentation.
- Conclusion: With minimal event reconstruction, one CVN algorithm achieved excellent signal-background separation in both NOvA oscillation channels.The channels are electron-neutrino appearance and muon-neutrino disappearance.
- Conclusion: CVN demonstrates that CNNs can work effectively on non-natural images formed from sampling-calorimeter readouts.The authors expect the approach to transfer to multiple detector technologies and analyses.
- Future work: The same training and architecture can be extended to ν NC and ντ CC identification for sterile-neutrino and exotic-physics searches.These interaction classes are already used by NOvA in such searches.
- Future work: Initial particle-identification studies with isolated particle-cluster images showed improved efficiency and purity over other NOvA likelihood methods.The authors are also exploring semantic segmentation to identify particles contributing to each detector-readout pixel.