Enhancing Event Reconstruction in Hyper-Kamiokande with Machine Learning: A ResNet Implementation

This paper demonstrates that a ResNet-based machine learning framework (WatChMaL) can achieve particle classification and kinematic reconstruction for Hyper-Kamiokande events with accuracy comparable to traditional methods while offering a massive speed-up of over 30,000 times, thereby enabling the processing of the experiment's required large-scale Monte Carlo datasets.

The Gist

Imagine a giant, underwater camera the size of a skyscraper, filled with pure water and lined with thousands of sensitive eyes (photomultiplier tubes). This is Hyper-Kamiokande, a next-generation experiment in Japan designed to catch ghostly particles called neutrinos.

When a neutrino bumps into an atom in the water, it creates a flash of blue light (Cherenkov radiation) that looks like a ring on the walls of the tank. Scientists need to analyze these rings to figure out:

1. What kind of particle made the ring (an electron, a muon, a photon, or a pion?).
2. Where the crash happened.
3. Which way the particle was flying.
4. How fast it was going.

The Problem: The "Slow Cook"

Currently, scientists use a traditional method (called fiTQun) to analyze these rings. Think of this like a master chef hand-crafting a single, perfect meal. The chef (the algorithm) tastes every single ingredient, measures every spice, and calculates the perfect recipe to understand the dish.

The problem? The chef is incredibly slow. It takes about 60 seconds to analyze just one event.

To get precise results about the universe, Hyper-Kamiokande needs to simulate and analyze millions of these events, including millions of "what-if" scenarios to account for measurement errors. If they use the slow chef, it would take centuries to process all the data needed for their experiments. It's like trying to feed a stadium by cooking one burger at a time.

The Solution: The "AI Super-Scanner"

This paper introduces a new approach using Machine Learning, specifically a type of AI called ResNet (a Residual Network).

Instead of a chef tasting every ingredient, imagine a super-fast security scanner at an airport. You don't need to know the exact recipe of the luggage to know if it's safe or dangerous; you just need to recognize the pattern.

Here is how the new system works:

1. Turning Light into Pictures: The raw data from the water tank (charges and timing of light hits) is flattened out and turned into a 2D image, like a photograph of the light ring.
2. Training the AI: The researchers fed the AI millions of these "photos" from computer simulations. They taught it to look at the picture and instantly say: "That's a muon," or "That's an electron," and then immediately guess the speed and direction.
3. The Result: The AI learned to recognize the subtle shapes and patterns of the light rings just like a human expert, but without the slow, step-by-step calculation.

The Magic Numbers

The results are staggering:

• Speed: The AI can analyze one event in 1 to 2 milliseconds. That's roughly 30,000 to 50,000 times faster than the traditional method. It's the difference between a snail and a supersonic jet.
• Accuracy: Despite being incredibly fast, the AI is just as good as the slow chef. It gets the speed, direction, and location of the particles almost exactly right.
• Superpowers: The AI is actually better at telling the difference between tricky particles (like an electron vs. a photon) than the old method, which has struggled with this for decades.

Why This Matters

Think of the traditional method as a hand-written diary and the new AI method as a high-speed search engine.

• The Old Way: If you want to find a specific fact in a library of a million books, you have to read every page of every book. It's accurate, but impossible to finish.
• The New Way: The AI scans the whole library in a blink, finding the exact pages you need instantly.

Because the AI is so fast, scientists can now run the massive, complex simulations they need to understand neutrino oscillations (how neutrinos change flavors) and the matter-antimatter imbalance in the universe. These are questions that could explain why we exist at all.

The Bottom Line

This paper proves that we can swap the "slow, perfect chef" for a "fast, super-smart AI scanner" without losing the quality of the meal. This breakthrough means that the Hyper-Kamiokande experiment can finally process the massive amounts of data it needs to unlock the secrets of the universe, turning a task that would take a lifetime into one that takes a weekend.

Technical Summary

1. Problem Statement

The Hyper-Kamiokande (HK) experiment, a next-generation water Cherenkov detector, aims to achieve unprecedented precision in measuring neutrino oscillation parameters, particularly the CP-violating phase $\delta_{CP}$. Achieving these goals requires processing massive Monte Carlo (MC) datasets with systematic variations to control uncertainties.

• Computational Bottleneck: Traditional event reconstruction relies on maximum-likelihood algorithms (e.g., fiTQun). While accurate, these methods have a per-event runtime of approximately minutes ($O(\text{min})$). Processing the vast, systematically varied MC samples required for HK is computationally prohibitive.
• Topological Challenges: Likelihood-based methods struggle with complex event topologies, such as overlapping Cherenkov rings (from $\pi^0 \to \gamma\gamma$ decays), vertices near detector walls, and events with sparse photon statistics.
• Need for Scalability: There is an urgent need for a reconstruction method that is both scalable (capable of handling MC-scale production) and robust against difficult topologies, without sacrificing the resolution required for precision physics.

2. Methodology

The authors propose a deep learning-based reconstruction framework using Residual Networks (ResNets) implemented within the WatChMaL (Water Cherenkov Machine Learning) framework.

Data Generation and Input Representation

• Simulation: Single-particle events (electrons $e$, muons $\mu$, gamma rays $\gamma$, and neutral pions $\pi^0$) were simulated using WCSim. Kinetic energies ranged from the Cherenkov threshold up to 2 GeV.
• Input Format: The detector's cylindrical surface (instrumented with ~20,000 50 cm PMTs) was "unwrapped" into a $190 \times 189$ 2D image.
  • Channels: Each pixel contains two channels:
    1. Hit Time: Time relative to the readout window start ($T_0$).
    2. Integrated Charge: Total photo-electrons collected within the 1.35 $\mu$s window.
  • Preprocessing: The image was rotated by $45^\circ$ to align PMTs into a regular grid, optimizing convolutional kernel efficiency.

Network Architectures

Seven distinct ResNet-152 models were developed:

1. Regression Models (6 total):
   • Three models for electrons: predicting interaction vertex ($x,y,z$), direction ($\hat{u}$), and total energy ($E_{total}$).
   • Three models for muons: predicting the same kinematic variables.
   • Loss Functions: Huber loss for vertex/direction; Relative Huber loss for energy (to prevent high-energy events from dominating the error metric).
2. Classification Model (1 total):
   • A four-class classifier ($e, \mu, \gamma, \pi^0$) using cross-entropy loss and Softmax output.

Training Strategy

• Dataset: ~$10^7$ events per particle species, split 96:2:2 (Train:Val:Test).
• Hardware: Training performed on four NVIDIA A100 GPUs.
• Selection Criteria: Evaluation was restricted to "fully contained" events with sufficient signal (>200 photo-electrons) to ensure fair comparison with traditional methods.

3. Key Contributions

• First Application to HK Far Detector: This is the first study to apply deep learning to the full Hyper-Kamiokande far detector geometry (as opposed to the Intermediate Water Cherenkov Detector).
• Direct $e$-$\gamma$ Separation: The model successfully distinguishes between electrons and single photons ($e$ vs. $\gamma$), a task previously considered unsolved in water Cherenkov detectors (even for Super-Kamiokande's likelihood methods).
• Scalable Pipeline: The work demonstrates a viable path to replacing or augmenting likelihood-based reconstruction for large-scale MC production, addressing the primary computational bottleneck of the HK experiment.

4. Key Results

Kinematic Reconstruction (Regression)

The ResNet models achieved resolutions broadly consistent with, and in some cases superior to, traditional fiTQun methods (based on indirect comparison with Super-Kamiokande-IV data):

• Momentum Resolution:
  • Muons: 1.35% (vs. ~2.26% for fiTQun).
  • Electrons: 2.39% (vs. ~2.90% for fiTQun).
• Vertex Resolution:
  • Muons: 28.2 cm.
  • Electrons: 25.4 cm.
• Angular Resolution:
  • Muons: 1.25$^\circ$.
  • Electrons: 1.94$^\circ$.
• Bias: The models exhibited low fractional bias (<0.25% for muons, <0.5% for electrons) across the momentum range.

Particle Identification (Classification)

The classifier demonstrated high separability, measured by the Area Under the ROC Curve (AUC):

• $e$ vs. $\mu$: AUC = 0.9999992 (Near-perfect separation).
• $e$ vs. $\pi^0$: AUC = 0.9526.
• $e$ vs. $\gamma$: AUC = 0.633.
  • Significance: While 0.633 seems modest, it represents the first successful statistical separation of electrons from single photons in this context, outperforming the indicative fiTQun baseline.

Computational Performance (Speed)

The most significant finding is the massive reduction in inference time:

• ResNet Inference: 1.3 ms (muons) and 1.7 ms (electrons) per event on a single GPU.
• fiTQun Inference: ~67 s (muons) and ~54 s (electrons) per event on CPU.
• Speed-up Factor: $3.2 \times 10^4$ to $5.2 \times 10^4$.

5. Significance and Implications

• Feasibility of Precision Analyses: The orders-of-magnitude speed-up makes it computationally feasible to generate the massive, systematically varied MC samples required for Hyper-Kamiokande's precision oscillation analyses (e.g., $\nu_\mu \to \nu_e$ appearance and $\delta_{CP}$ sensitivity).
• Robustness: The neural networks maintained performance in "difficult" regions where likelihood fits typically degrade, such as events near detector walls or with low photon statistics. This suggests ML can reduce the need for aggressive fiducial volume cuts.
• Future Outlook: While traditional algorithms (fiTQun) will remain vital for calibration and cross-checks, this work establishes deep learning as a scalable alternative for the bulk of event reconstruction. Future work will focus on extending these methods to complex multi-ring topologies (e.g., full $\pi^0$ reconstruction) and integrating systematic uncertainties directly into the training.

In conclusion, this paper validates that ResNet-based deep learning can match the reconstruction accuracy of state-of-the-art likelihood methods while reducing computational costs by four to five orders of magnitude, thereby unlocking the full potential of the Hyper-Kamiokande experiment.