LArTPC hit-based topology classification with quantum machine learning and symmetry

This paper presents a quantum machine learning approach using rotationally symmetric quanvolutional neural networks to classify track-like and shower-like topologies in LArTPC neutrino events, finding that while these quantum models outperform similarly sized classical networks, they are ultimately surpassed by classical models with significantly more parameters.

The Gist

Imagine you are a detective trying to solve a mystery inside a giant, frozen room filled with invisible gas. This room is a Liquid Argon Time Projection Chamber (LArTPC), a high-tech detector used to catch neutrinos—ghostly particles that pass through everything.

When a neutrino bumps into an atom in this room, it creates a splash of energy. This splash leaves a trail, like a footprint in the snow. But here's the problem: sometimes the trail is a straight, thin line (like a track from a muon), and sometimes it's a messy, fluffy cloud (like a shower from a photon).

To understand the mystery, scientists need to instantly tell the difference between a "straight line" and a "fluffy cloud" for every single pixel in the photo of the event. This is called topology classification.

The Old Way vs. The New Way

The Old Way (Classical AI):
Imagine you have a team of human detectives looking at these photos. They use standard rules to spot lines and clouds. In the world of computers, this is done by Classical Neural Networks (like the AI that recognizes cats in your photos). They are very good, but to get them to be really good, you need to hire a massive team (millions of parameters) and give them a huge library of examples to study.

The New Way (Quantum Machine Learning):
The authors of this paper asked: "What if we used a Quantum Computer to help?"
Think of a classical computer as a super-fast calculator that checks one thing at a time. A quantum computer is like a magical crystal ball that can look at many possibilities at once.

They built a new type of AI called a Quanvolutional Neural Network.

• The Analogy: Imagine you are looking at a small patch of the photo (a "window"). Instead of just using a standard magnifying glass (a classical filter), you put that patch through a quantum kaleidoscope. This kaleidoscope twists and turns the information in a way that classical computers can't easily do, revealing hidden patterns.

The Experiment: Two Datasets

The team tested their "Quantum Kaleidoscope" on two types of evidence:

1. The MicroBooNE Dataset: Real data from a real experiment. It's like looking at a messy crime scene with thousands of footprints, some overlapping, some faint.
2. The "Neutrino-Like" Dataset: A custom-made simulation. Imagine a scenario where a straight line and a fluffy cloud start from the exact same spot.
   • Easy Mode: They start far apart (easy to tell apart).
   • Hard Mode: They start right on top of each other, overlapping so much it looks like a single blob (very hard to tell apart).

What Did They Find?

Here is the verdict, translated into plain English:

1. The Quantum Team is Efficient:
When the Quantum AI and the Classical AI were given the same amount of "brain power" (the same number of adjustable settings), the Quantum AI won. It was better at spotting the difference between a track and a shower. It's like a small, highly trained quantum detective solving the case faster than a regular detective with the same resources.

2. The Classical Team Wins with Size:
However, if you give the Classical AI a massive team (100 times more brain power), it beats the Quantum AI. The Quantum AI is great at being efficient, but the Classical AI is a brute-force champion. If you throw enough money and computing power at the classical model, it wins.

3. The "Symmetry" Twist:
The researchers tried to teach the AI that "up is the same as down" or "left is the same as right" (rotational symmetry). They thought this would help the AI understand that a track is a track, no matter how it's rotated.

• The Result: It didn't help much. In fact, sometimes it made things slightly worse. It's like trying to teach a dog that a ball is a ball whether it's red or blue; the dog already figured that out on its own, and the extra rules just confused it.

4. The "Patch Size" Problem:
The AI looks at the image in small squares (patches).

• If the square is too small, the AI sees a dot and doesn't know if it's a line or a cloud.
• If the square is too big, the AI gets overwhelmed by too much noise.
• Sweet Spot: They found that a medium-sized square (21x21 pixels) was the perfect size for the models to work.

Why Does This Matter?

We are building the next generation of neutrino detectors (like DUNE), which will be massive. They will produce so much data that our current computers might get overwhelmed.

This paper shows that Quantum Machine Learning is a promising tool. Even though we can't run these models on real quantum computers yet (they are too small and noisy), the math of how they work suggests they are incredibly efficient.

The Bottom Line:
Think of this as a prototype for a hybrid car. Right now, the "gas engine" (Classical AI) is still the most powerful, but the "electric motor" (Quantum AI) is surprisingly efficient and might be the key to solving the most difficult, overlapping mysteries in the future when we have bigger detectors and need to be smarter, not just bigger.

The authors conclude that while Quantum AI isn't ready to take over the whole job yet, it's a powerful new tool in the toolbox that could help us understand the universe's most elusive particles.

Technical Summary

1. Problem Statement

In Liquid Argon Time Projection Chamber (LArTPC) experiments (such as MicroBooNE and the future DUNE), accurate reconstruction of neutrino events is critical. These detectors produce 2-D images of particle interactions where charged particles leave distinct signatures:

• Tracks: Linear structures (e.g., muons, protons).
• Showers: Dense, branching collections of pixels (e.g., electrons, photons, positrons).

The Challenge: Separating track-like and shower-like topologies at the "hit" (pixel) level is a fundamental but difficult reconstruction task. Current machine learning (ML) methods, while effective, face challenges in dense regions where tracks and showers overlap. Furthermore, there is a need to explore Quantum Machine Learning (QML) to see if it offers advantages in feature extraction or parameter efficiency for these high-energy physics (HEP) problems, particularly given the constraints of current Noisy Intermediate-Scale Quantum (NISQ) devices.

2. Methodology

Datasets

The study utilized two datasets to train and evaluate models:

1. MicroBooNE Open Dataset: ~750,000 simulated neutrino interaction events (Booster Neutrino Beam) overlaid with cosmic rays (cosmic rays were removed for this study to align with DUNE's low-background environment). Events are represented as three 2-D images corresponding to the detector's wire readout planes (U, V, Y).
2. Custom "Neutrino-like" Dataset: A synthetic dataset generated using LArSIMple (based on Geant4). It features events with one electron (shower) and one muon (track) originating from a common vertex with variable opening angles ($0^\circ$ to $180^\circ$). This allows for controlled testing of classification difficulty based on spatial overlap.

Approach: Patch-Based Classification

Instead of classifying entire events, the authors performed pixel-level classification.

• The 2-D event images were subdivided into $N \times N$ pixel patches.
• For each patch, the central pixel was labeled as either "track" (0) or "shower" (1) based on the dominant particle type in the simulation.
• The goal was to train models to predict the label of the central pixel using the surrounding context within the patch.

Model Architectures

The study compared classical Convolutional Neural Networks (CNNs) against Quantum-Enhanced architectures, specifically Quanvolutional Neural Networks.

• Classical Baselines:

  • CNN: A compact 2-layer network.
  • Deep CNN: A 6-layer network.
  • Deep GCNN: A deep CNN incorporating $C_4$ rotational symmetry (equivariant to 90-degree rotations).

• Quantum-Enhanced Models (Quanv):

  • Quanvolutional Layer: Replaces standard convolution filters with quantum circuits. A sliding window of pixels is embedded into a Parameterized Quantum Circuit (PQC) via a rotation embedding ($U_{F,W}$). The circuit processes the data, and an expectation value of an observable is returned as the new feature map value.
  • Symmetry Extensions: The authors extended the standard quanvolution to respect $C_4$ rotational symmetry (invariance/equivariance under 90-degree rotations), a novel application for LArTPC images.
  • Classifiers: Models used either a classical Multi-Layer Perceptron (MLP) or a Geometric Quantum Classifier (an invariant quantum circuit) for the final classification step.

Symmetry Implementation

The paper rigorously defines how quanvolutions inherit equivariance from classical convolutions. By using identical circuit structures for each group element (e.g., rotating the input and applying the same circuit logic), the models respect the symmetries of the Euclidean group (specifically $C_4$).

3. Key Contributions

1. Novel Application of QML in HEP: This is one of the first studies applying QML to hit-level topology classification in LArTPC experiments, moving beyond simple event classification.
2. Redefinition of Quanvolution: The authors correct a common misconception in the literature. They demonstrate that a quanvolution is not a "quantum filter" performing a dot product, but rather a quantum circuit acting as a non-linear transformation of a feature map that inherits equivariance properties from the underlying convolution structure.
3. Symmetry in Quantum Architectures: For the first time, the study extends quanvolutional networks to include symmetries beyond translation (specifically $C_4$ rotational symmetry), integrating geometric QML principles into the architecture.
4. Benchmarking Framework: The study provides a rigorous comparison of QML vs. Classical ML under controlled conditions (parameter counts, patch sizes, and data scarcity).

4. Results

Performance on MicroBooNE Data

• Parameter Efficiency: Quantum-enhanced models (Quanv) consistently outperformed their classical counterparts (CNN) when constrained to a similar number of trainable parameters.
• Scalability: However, classical models with significantly more parameters (approx. two orders of magnitude more, or $100\times$) outperformed the quantum models.
• Patch Size Sensitivity: Larger models (Deep CNN/Deep GCNN) benefited from larger patch sizes (up to 49 pixels), utilizing more context. Smaller models (including standard Quanv) saw performance degradation with very large patches due to noise. The optimal patch size identified was 21 pixels.
• Symmetry Impact: The inclusion of $C_4$ rotational symmetry showed no significant advantage for the quantum models. In fact, the non-symmetric Quanv often performed best. For deep classical models, symmetry offered a slight, consistent improvement.

Performance on Custom "Neutrino-like" Dataset

• Difficulty Scaling: The models were tested on opening angles ranging from $0^\circ$ (hard, high overlap) to $180^\circ$ (easy).
• Quantum Advantage in Complexity: The Quanvolutional network consistently outperformed the shallow CNN across all difficulty levels, with the largest performance gap occurring in the "Hard" category ($0^\circ-5^\circ$). This suggests QML is better at extracting features in dense, overlapping topologies.
• Classical Dominance: Despite the quantum advantage over shallow networks, the Deep CNN (with many more parameters) still outperformed the Quanv models across all difficulty levels.

Data Efficiency

• In low-data regimes (training on 100–1000 samples), Quanvolutional networks demonstrated superior learning capabilities compared to shallow CNNs, suggesting richer feature representations.

5. Significance and Future Outlook

• Hybrid Reconstruction: The results suggest that while current quantum hardware cannot yet replace large-scale classical deep learning, QML offers a competitive alternative for parameter-constrained scenarios or specific sub-tasks where classical models struggle with dense data.
• Future Experiments (DUNE): The study highlights the potential for QML in the Deep Underground Neutrino Experiment (DUNE). Given the massive complexity of DUNE data, a hybrid approach (using QML for specific, difficult sub-routines within the Pandora reconstruction framework) could be valuable.
• Theoretical Insight: The paper clarifies the theoretical relationship between convolutions and quanvolutions, providing a roadmap for designing symmetry-aware quantum circuits in high-energy physics.
• Limitations: The study is limited by the computational cost of simulating quantum circuits on classical hardware, restricting the training data size and the depth of the quantum circuits. Future work requires leveraging GPU acceleration and more efficient simulation techniques to scale these models to real-world LArTPC data volumes.

Conclusion: The paper demonstrates that quantum-enhanced models can achieve superior performance to classical models with comparable parameter counts, particularly in dense, complex topological regions. However, they are currently outperformed by large-scale classical deep learning. The inclusion of rotational symmetry in quantum models did not yield a clear benefit in this specific context, indicating a need for further exploration of different symmetry groups.