Deep-learning-based low-energy trigger algorithms for the Hyper-Kamiokande experiment

This paper demonstrates that deep-learning-based trigger algorithms, particularly a supervised neural network and an MPDR-based anomaly detection model, significantly outperform traditional hit-count triggers in identifying low-energy neutrino events for the Hyper-Kamiokande experiment while maintaining real-time feasibility with sub-millisecond GPU inference latencies.

The Gist

Imagine the Hyper-Kamiokande experiment as a massive, ultra-sensitive underwater listening station. Its job is to "hear" tiny ripples caused by ghostly particles called neutrinos. However, this ocean is incredibly noisy. The detector is constantly bombarded by random static and background chatter (detector noise), making it very hard to spot the faint, specific whispers of the neutrinos we are looking for, especially the low-energy ones.

The paper presents a new way to filter this noise using Artificial Intelligence (AI), acting like a super-smart security guard who can instantly decide whether to save a recording or ignore it.

Here is a breakdown of their approach using everyday analogies:

1. The Problem: Finding a Whisper in a Storm

In the past, the detector used a simple rule to decide what to save: "If we hear this many clicks from our sensors, save it." This is like a bouncer at a club who only lets people in if they are shouting.

• The Flaw: Low-energy neutrinos are quiet. They don't make enough "clicks" to trigger the old rule, so they get ignored. Meanwhile, the random noise sometimes makes enough clicks to trick the system, wasting storage space on garbage data.

2. The Solution: The AI "Pattern Detective"

The researchers trained three different types of AI "detectives" to look at the data. Instead of just counting clicks, these detectives look at the shape, timing, and location of the signals, much like a detective looking for a specific fingerprint rather than just counting how many people are in a room.

Detective A: The Supervised Teacher (The "Signal Hunter")

• How it works: This AI was shown millions of examples of "real neutrino whispers" and "fake noise static." It learned exactly what a real signal looks like.
• The Trick: It uses a sophisticated brain architecture (called a Transformer) that understands how different sensors talk to each other. It doesn't just look at one sensor; it sees the whole "dance" of the particles.
• The Result: It is incredibly good at spotting the quiet whispers. For a very faint signal (3 MeV), it caught 76.7% of them, whereas the old "count the clicks" method only caught 26.4%. It's like upgrading from a metal detector that only finds big coins to one that finds tiny gold flakes.

Detective B: The Noise Specialist (The "Anomaly Hunter")

• How it works: This AI was only shown the background noise. It learned to memorize what "normal static" looks like perfectly.
• The Trick: When it sees something that doesn't quite fit the "noise pattern" (even if it doesn't know exactly what the signal is), it flags it as "suspicious." This is called Anomaly Detection.
• The Result: One version of this (called MPDR) was surprisingly good, catching 31.8% of the signals. It's like a security guard who knows the sound of the wind so well that if a door creaks slightly differently, they know something is up, even if they don't know what the intruder looks like.

3. The "Magic" of Speed

Usually, fancy AI is slow and requires massive computers. The researchers tested these detectives on powerful graphics cards (GPUs) and found they could make a decision in less than a millisecond.

• The Analogy: Imagine a security guard who can scan a thousand people in the time it takes to blink. This speed means they can be used in real-time, filtering data as it happens rather than waiting to analyze it later.

4. What They Found

• The Winner: The "Signal Hunter" (Supervised AI) was the best at finding the neutrinos, especially the faint ones.
• The Runner-Up: The "Anomaly Hunter" (MPDR) was also very good and has a special advantage: it doesn't need to know what the signal looks like beforehand. It just needs to know what the noise doesn't look like. This is great because if our understanding of the neutrinos changes, this AI still works.
• The Loser: A simple "count the clicks" method (the old way) missed most of the low-energy signals.
• Bonus: They also tested if these AI could spot "gamma rays" (a different type of particle signal). The AI was much better at this than the old method too.

Summary

The paper proves that by using modern AI to look at the patterns of light and time in the detector, rather than just counting how many sensors went off, we can hear the "whispers" of the universe that were previously too quiet to detect. This allows scientists to push the boundaries of what they can see, potentially revealing secrets about the sun, exploding stars, and the fundamental laws of physics.

Technical Summary

Technical Summary: Deep-Learning-Based Low-Energy Trigger Algorithms for the Hyper-Kamiokande Experiment

Problem Statement
The Hyper-Kamiokande (Hyper-K) experiment aims to conduct precision measurements of low-energy neutrinos (below 7 MeV), including solar, supernova, and reactor neutrinos, as well as the Diffuse Supernova Neutrino Background (DSNB). A critical bottleneck for these studies is the Data Acquisition (DAQ) system's ability to distinguish faint Cherenkov signals from dominant detector noise in real-time. Traditional triggers based on simple hit-count thresholds (NHits) suffer from low efficiency at these energies because the number of hits from low-energy signals significantly overlaps with the distribution of dark noise hits. To fully exploit Hyper-K's physics potential, particularly for resolving tensions in $\Delta m^2_{21}$ measurements and detecting rare supernova events, the detection threshold must be pushed below 7 MeV. This requires trigger algorithms capable of recognizing complex spatiotemporal patterns in sparse PMT hit data while operating under stringent runtime constraints (sub-millisecond latency).

Methodology
The authors propose and evaluate two distinct machine-learning approaches for low-energy triggering, both operating on sparse point-cloud representations of Photomultiplier Tube (PMT) hits (position, time, and charge) within 400 ns windows:

1. Supervised Transformer Classifiers:

   • Architecture: A 6-layer Transformer encoder utilizing self-attention mechanisms to model global event context and local hit correlations.
   • Input Representation: Hits are encoded using cylindrical coordinates aligned with the detector geometry. Features include spatial position (sin/cos of azimuth, normalized radius/height), hit time, and integrated charge.
   • Training Strategies: Two supervision levels were tested:
     • Event-level: Predicting a single signal/noise label for the entire event.
     • Hit-level: Predicting the signal/noise status of individual hits, with the final event score derived from the maximum hit probability.
   • Data: Trained on 1 million simulated single-electron events (0–7 MeV) mixed with detector noise.

2. Unsupervised Anomaly Detection (Noise-Only Learning):

   • Approach: Models trained exclusively on detector noise to identify events that deviate from the learned noise distribution.
   • Autoencoder: A transformer-based autoencoder (Encoder-Decoder) trained to reconstruct noise events. It uses a unit-norm latent space constraint and a fixed number of candidate hits with existence probabilities to handle variable hit multiplicities.
   • Manifold Projection–Diffusion Recovery (MPDR): An energy-based model (specifically MPDR-S) that learns a scalar energy function $E_\theta(X)$. It employs a two-phase training process:
     • Phase I: Train the autoencoder on noise.
     • Phase II: Freeze the autoencoder and train an energy function to distinguish true noise from "hard" synthetic negatives. These negatives are generated by perturbing noise in the latent space and partially recovering them via a Langevin-style MCMC process.
   • Input: Uses only spatial (unit sphere parametrization) and temporal information; charge is omitted for these models.

Key Contributions and Results
The study provides a comprehensive benchmark of these algorithms against a traditional NHits trigger and evaluates their performance in terms of signal efficiency, discrimination power, and runtime.

• Signal Identification Efficiency:

  • The supervised hit-level classifier achieved the highest performance, identifying 76.7% of 3 MeV electrons at a false trigger rate of <10 kHz. This significantly outperforms the traditional NHits trigger (26.4%) and the MPDR approach (31.8%).
  • The supervised event-level classifier performed similarly above 4 MeV but showed a 5–7% efficiency drop compared to the hit-level version in the 2–3 MeV range.
  • The MPDR approach outperformed the NHits trigger across the energy spectrum (e.g., 31.9% vs. 26.4% at 3 MeV), demonstrating that anomaly detection can improve low-energy triggers without explicit signal modeling.
  • The pure autoencoder (without MPDR) performed worse than the NHits trigger (AUROC 0.76 vs. 0.86), indicating that reconstruction error alone is insufficient for this task.

• Feature Importance:

  • Feature ablation studies revealed that hit time is the most critical feature. Removing charge information had negligible impact on performance (AUROC remained ~0.90), whereas removing time information caused a drastic performance drop. This suggests that spatio-temporal correlations are the primary discriminators, potentially reducing reliance on precise PMT charge calibration.

• Neutron Tagging and Gamma Identification:

  • Models trained on electrons were applied to 2.2 MeV gamma rays (simulating neutron capture on hydrogen). The supervised hit-level classifier achieved a 25.4% efficiency, compared to 4.2% for NHits. The MPDR achieved 4.8%.

• Runtime Performance:

  • Inference benchmarks on an NVIDIA A100 GPU showed that all models operate well below the millisecond scale.
  • The MPDR energy network demonstrated the highest throughput (30,000 windows/s), followed by the autoencoder (20,000 windows/s).
  • The supervised transformer, while powerful, had higher latency and memory usage due to the computational cost of calculating pairwise relative positional biases for attention.

Significance and Claims
The paper claims that deep-learning-based triggers offer a viable path to lowering the detection threshold for Hyper-Kamiokande below 7 MeV, a regime where traditional triggers fail due to noise dominance.

• Feasibility of Real-Time ML: The study demonstrates that complex deep learning models, including Transformers and energy-based models, can be integrated into online DAQ systems with inference latencies in the tens of microseconds, satisfying the strict timing requirements of the experiment.
• Robustness via Anomaly Detection: The MPDR approach is highlighted as a significant proof-of-concept for anomaly detection in particle physics. Its ability to improve efficiency without requiring detailed signal modeling makes it robust against potential mismodeling of signal physics, provided the detector noise is well-understood.
• Complementarity: The results suggest that combining traditional hit-count triggers with machine-learning algorithms (e.g., using NHits as a pre-trigger) could optimize resource usage while maximizing signal retention.
• Physics Impact: By enabling the efficient acquisition of data below 7 MeV, these algorithms directly support critical physics goals, including the resolution of solar neutrino oscillation parameters, the detection of the DSNB, and the study of supernova dynamics.

The authors conclude that while the supervised hit-level classifier currently offers the best performance, the MPDR approach provides a robust, signal-agnostic alternative. They note that further gains could be realized by dedicated models incorporating temporal correlations between prompt lepton signals and delayed gamma rays for neutron tagging, which was not fully explored in this specific study.