Enhanced Ionization Charge Identification in the Short-Baseline Neutrino Program Neutrino Detectors with Deep Neural Networks

This paper introduces a deep neural network-based region of interest detection method (DNN ROI) for the Short-Baseline Neutrino Program's liquid argon time projection chambers, which outperforms traditional thresholding algorithms in identifying ionization charge and reconstructing neutrino interactions while demonstrating greater robustness against detector variations.

The Gist

Imagine you are trying to listen to a specific conversation in a very loud, crowded room. The room is filled with static, echoes, and people shouting over each other. This is essentially what scientists face when they try to detect neutrinos—tiny, ghost-like particles that barely interact with anything.

The paper describes a new "smart ear" (a Deep Neural Network, or DNN) designed to help two specific detectors, SBND and ICARUS, hear these ghostly conversations clearly. Here is how it works, broken down into simple concepts:

1. The Problem: The "Static" in the Room

The detectors used in this experiment are like giant 3D cameras filled with liquid argon. When a neutrino hits an atom, it creates a trail of electrons (like a spark). The detector tries to record these sparks as they drift toward wires.

However, the recording is messy:

• Noise: There is electronic static (like radio static) that drowns out the signal.
• The "Double-Edge" Sword: On some wires, the signal looks like a wave that goes up and then immediately down, canceling itself out. It's like trying to see a shadow that keeps flipping between light and dark, making it hard to tell where the object actually is.
• Old Method: The traditional way to find these sparks was like using a simple "volume knob." If the sound got louder than a certain level, the computer assumed it was a signal. If it was too quiet, it was ignored. This worked for loud, clear sounds (like a shout) but failed for complex, quiet, or "cancelled out" whispers.

2. The Solution: The "Smart Detective" (DNN ROI)

The authors built a new system called DNN ROI (Region of Interest). Instead of just listening for loud noises, this system acts like a super-smart detective that looks at the whole picture at once.

• Looking at the Whole Room: Instead of checking one wire at a time, the AI looks at a 2D image of the entire detector. It sees how the wires interact with each other.
• Cross-Checking Clues: The detector has three layers of wires. The AI checks if a "spark" appears in the same spot on all three layers at the same time. If it does, it's almost certainly a real particle. If it only appears on one layer, it's likely just static noise.
• Learning from Mistakes: The AI was trained on millions of simulated events. To make it tougher, the scientists "tricked" the AI during training by randomly turning off wires or adding extra static. This is like training a detective by putting them in a room where the lights flicker and some microphones are broken, so they learn to find the truth even when things go wrong.

3. The Results: A Clearer Picture

When they tested this new AI against the old "volume knob" method, the results were impressive:

• Finding the Hard-to-See: The AI was much better at finding long, thin tracks of particles that were almost parallel to the wires (which usually get lost in the "cancellation" effect). It was also better at spotting "showers" of particles (like a spray of sparks from a single hit).
• Measuring Energy: Because the AI found more of the signal and ignored more of the noise, the scientists could measure the energy of the particles much more accurately. It's like the difference between guessing the weight of a package by looking at a blurry photo versus weighing it on a precise scale.
• Robustness: Even when the detector had "glitches" (like dead wires or extra noise), the AI didn't get confused. It knew to ignore the broken parts and focus on the working ones. The old method, however, would often get tripped up by these glitches.

4. Why It Matters

This isn't just about making better pictures; it's about physics. By cleaning up the data so effectively, the scientists can now study the properties of neutrinos with much higher precision.

The paper concludes that this "smart detective" is now being used for real data in the SBND and ICARUS experiments. It's a flexible tool that can adapt to different detector conditions, ensuring that the scientists don't miss any of the subtle clues these ghostly particles leave behind. The authors also note that this same technology could be adapted for future, even larger experiments (like DUNE) to help them see the universe more clearly.

Technical Summary

Technical Summary: Enhanced Ionization Charge Identification in the Short-Baseline Neutrino Program

Problem Statement
The Short-Baseline Neutrino (SBN) Program at Fermi National Accelerator Laboratory (FNAL) utilizes Liquid Argon Time Projection Chambers (LArTPCs), specifically the SBND and ICARUS detectors, to reconstruct charged particle trajectories with high spatial resolution. A critical early step in event reconstruction is signal processing, which involves identifying Regions of Interest (ROIs) containing true ionization signals within raw waveforms. Traditional ROI detection relies on wire-by-wire thresholding and heuristics based on particle trajectory connectivity. While effective for sparse, high-energy tracks, these methods struggle with complex charge depositions, such as extended trajectories perpendicular to wire planes (producing prolonged, non-Gaussian signals due to bipolar induction effects) and electromagnetic showers. Furthermore, traditional algorithms can be sensitive to detector performance variations, including noise fluctuations, electron lifetime changes, and wire plane intransparency.

Methodology
To address these limitations, the authors implement a Deep Neural Network (DNN) based ROI detection method (DNN ROI), originally introduced in Ref. [8], adapted for the SBND and ICARUS detectors. The approach frames ROI detection as a 2D semantic segmentation task, labeling each pixel in the detector readout as either signal or noise.

• Input Pre-processing: The network ingests three two-dimensional image channels derived from deconvolved waveforms:
  1. ROI Filter Output: A Wiener-like filtered waveform optimized for signal-to-noise ratio.
  2. Two-Plane (MP2) Coincidence: A binary map identifying channels where activity coincides across at least two wire planes within a common time window.
  3. Three-Plane (MP3) Coincidence: A binary map highlighting channels with simultaneous signals across all three wire planes.
• Network Architecture: The system employs a U-ResNet architecture, combining the U-Net encoder-decoder structure with ResNet residual blocks. The network is optimized for computational efficiency via "chunking" (splitting images into smaller arrays) and downsampling (averaging over fixed tick intervals) to allow inference on CPUs within the LArSoft framework.
• Training and Augmentation: Training samples are generated using Monte Carlo simulations (GENIE, CORSIKA, GEANT4, and WireCell) covering BNB and NuMI neutrino interactions, cosmic rays, and specific challenging topologies like prolonged tracks and $\nu_e$ showers. To ensure robustness against real-world detector variations, the authors employ data augmentation strategies:
  • ICARUS: "OmniDetector" samples where simulation parameters (noise scales, electron lifetime, gain, signal shapes) are randomly varied to reflect observed detector instabilities.
  • SBND: Direct augmentation of ROI filter output images to simulate waveform smearing, pixel scaling, and masked (dead) wire bands.
  • Class Imbalance: A weighted binary cross-entropy loss function is used to address the sparse nature of signal pixels.

Key Contributions

1. Implementation of DNN ROI: The successful adaptation and deployment of a deep learning-based ROI finder for the SBND and ICARUS detectors, replacing traditional thresholding algorithms.
2. Robustness via Augmentation: A systematic study demonstrating that training with augmented samples (simulating dead wires, noise variations, and signal shape changes) significantly improves the network's resilience to detector defects and performance drifts.
3. Cross-Plane Integration: The utilization of geometric constraints (MP2 and MP3 coincidence) as input channels, which enhances the network's ability to distinguish true physics signals from noise and induction-plane artifacts.
4. Scientific Robustness Case Study: The work serves as a practical demonstration of "scientific robustness" in machine learning, showing that neural networks can provide unbiased results even when faced with data deformations typical of real experiments.

Results
The evaluation compares DNN ROI against traditional methods using both low-level metrics (pixel/ROI efficiency and purity) and high-level reconstruction metrics (charge extraction and shower completeness).

• Performance Improvement: DNN ROI outperforms traditional thresholding in both efficiency and purity across various event topologies. Improvements are particularly notable for:
  • Prolonged Tracks: Tracks at shallow angles to the drift field, where bipolar cancellation traditionally obscures signals.
  • Electromagnetic Showers: Complex topologies where traditional methods suffer from energy-dependent biases.
• Robustness to Variations:
  • In ICARUS, DNN ROI models trained on "OmniDetector" samples maintained stable performance across extreme variations (e.g., 20% increased noise, low electron lifetime), whereas the traditional algorithm showed significant degradation (up to 7% drop in Efficiency $\times$ Purity).
  • In SBND, networks trained with augmentation successfully identified and ignored dead wire regions, whereas models trained on nominal data produced unphysical ROIs in those areas.
• Real Data Validation: Analysis of SBND cosmic data confirms that DNN ROI preserves inter-plane charge balance comparable to the traditional method, validating its applicability beyond simulation.
• Ablation Studies: Removing input channels (MP2 or MP3) resulted in minor performance drops for general neutrino events but noticeable degradation for specific challenging topologies (prolonged tracks and showers), confirming the value of cross-plane geometric constraints.

Significance
The paper claims that DNN ROI provides a flexible and robust framework for signal processing in LArTPCs, overcoming the limitations of heuristic-based methods. By leveraging full 2D detector readout and cross-plane matching, the method improves the fidelity of ionization charge extraction and particle energy reconstruction. The authors emphasize that the demonstrated robustness against detector variations makes this approach suitable for current SBN operations and adaptable for future large-scale experiments, such as the Deep Underground Neutrino Experiment (DUNE). The work establishes a precedent for integrating machine learning into the core signal processing chains of neutrino experiments to handle complex detector conditions and improve physics reach.