Abstract
We present a deep learning approach for vertex reconstruction of neutrino-nucleus interaction events, a problem in the domain of high energy physics. In this approach, we combine both energy and timing data that are collected in the MINERvA detector to perform classification and regression tasks. We show that the resulting network achieves higher accuracy than previous results while requiring a smaller model size and less training time. In particular, the proposed model outperforms the state-of-the-art by 4.00% on classification accuracy. For the regression task, our model achieves 0.9919 on the coefficient of determination, higher than the previous work (0.96).
Introduction
Neutrinos are fundamental particles that rarely interact with matter, making them notoriously difficult to study. In experiments like MINERvA at Fermilab, physicists observe neutrino collisions with various nuclear targets (e.g., carbon, iron, lead) to understand these interactions. A critical step in this analysis is vertex reconstruction: identifying the precise 3D location where the neutrino hit a nucleus. Accurate reconstruction is essential for determining which target material was involved.
Traditional methods often struggle with complex event topologies, and previous deep learning models typically analyzed energy or timing data in isolation. This research addresses these limitations by developing a high-performance model that fuses both data types into a compact, efficient architecture.
Methodology
The researchers developed a multi-task deep learning model designed to fuse disparate data types into a unified spatial representation.
Data Representation and Fusion
The model treats detector data as multi-channel images. Specifically, it utilizes a 6-channel input consisting of three geometric views (X, U, and V orientations) for both energy (the magnitude of the interaction) and timing (the temporal sequence of sensor triggers).
Model Architecture
The architecture is inspired by ResNet and DenseNet, utilizing shortcut connections to facilitate gradient flow and feature concatenation to preserve spatial information across layers. The model is designed for two primary tasks:
- Classification: Identifying which of the 11 physical segments of the detector the interaction occurred in.
- Regression: Predicting the exact Z-coordinate (depth) of the vertex.
The team employed transfer learning for the regression task. After training the classification model, the convolutional layers were "frozen," and a new regression head was added to predict the precise numerical position based on the learned geometric features.
Dataset
The study used a dataset of approximately 1.45 million simulated neutrino-nucleus scattering events generated via the GENIE Monte Carlo generator and Geant4 toolkit. This high-fidelity simulation provides the ground-truth labels necessary for training, which are otherwise impossible to obtain from raw experimental data.
Results and Discussion
The proposed model significantly outperformed previous benchmarks. It achieved a 98.09% accuracy in segment classification, a 4% improvement over the previous state-of-the-art. In the regression task, the model reached a coefficient of determination ($R^2$) of 0.9919, compared to the previous best of 0.96.
Beyond accuracy, the model is highly efficient. By utilizing modern architectural techniques, the researchers reduced the model size to just 0.488 MB—30 times smaller than previous iterations (14.5 MB). Training time was also reduced by 75%, from 10 hours to 2.5 hours.
A key insight from this work is the complementarity of energy and timing data. While energy provides the primary signal, timing data helps resolve ambiguities in segments where energy patterns are similar, leading to the observed performance gains.
Conclusion
This work demonstrates that compact, multi-modal deep learning architectures can significantly improve vertex reconstruction in high-energy physics. Future research will focus on addressing data imbalances in rare event segments and exploring Recurrent Neural Networks (RNNs) to better capture the temporal dynamics of interactions. The ultimate goal is to apply these models to real-world experimental data from the MINERvA detector.
Summary updated