Enhancing Neutrinoless Double-Beta Decay Sensitivity of Liquid-Xenon Time Projection Chamber with Augmented Convolutional Neural Network

This paper proposes an augmented convolutional neural network (A-CNN) model that significantly enhances the sensitivity of liquid xenon time projection chambers to neutrinoless double-beta decay by achieving over 60% background rejection while maintaining 90% signal acceptance, thereby improving the projected sensitivity of experiments like XENONnT by approximately 40%.

The Gist

The Big Picture: Hunting for a Ghost in a Noisy Room

Imagine you are trying to hear a single, specific whisper in a stadium filled with thousands of people shouting, cheering, and clapping. That is essentially what physicists are doing when they search for Neutrinoless Double-Beta Decay (0νββ).

This is a hypothetical event where two neutrons in an atom turn into protons and emit electrons, but no anti-neutrinos. If we find this, it proves that neutrinos are their own antiparticles (like a mirror image of yourself that is actually you). This discovery would rewrite the rules of physics and explain why the universe is made of matter instead of being empty.

The problem? The "whisper" (the signal) is incredibly faint, and the "shouting" (background noise from radioactive materials in the detector) is deafening.

The Detective's Tool: The Liquid Xenon Chamber

The XENONnT experiment uses a giant tank of liquid xenon (a heavy, noble gas) as a detector. Think of this tank as a massive, ultra-sensitive swimming pool. When a particle hits the water, it creates a splash (light) and a ripple (electric charge).

The detector has cameras (called Photomultiplier Tubes or PMTs) all over the ceiling and floor to catch these splashes.

• The Signal (The Whisper): A neutrinoless decay happens in one tiny spot. It's a "single-site" event. Imagine a single raindrop hitting the pool.
• The Noise (The Shouting): Radioactive bits in the detector walls bounce around, hitting the water in multiple places. These are "multi-site" events. Imagine someone throwing a handful of pebbles into the pool; you see splashes in several different spots at once.

The Old Way vs. The New Way

The Old Way (Traditional Cuts):
Traditionally, scientists tried to filter out the noise by looking at the "shape" of the splashes. If the ripples were too spread out, they'd say, "That's a rock, not a raindrop," and throw it away. But this is like trying to sort marbles from sand by only looking at the color. It works okay, but you lose a lot of good marbles (signal) along with the sand (noise).

The New Way (The A-CNN):
The authors of this paper built a super-smart digital detective called an Augmented Convolutional Neural Network (A-CNN).

Think of the A-CNN as a music producer listening to a recording.

• The Input: Instead of just looking at the "shape" of the splash, the A-CNN listens to the entire waveform (the sound of the splash) in high definition. It hears the tiny vibrations, the timing, and the subtle echoes that human eyes miss.
• The "Augmentation" Trick: Here is the clever part. The A-CNN was trained on computer simulations, but real life is messy. To make the detective robust, the scientists "augmented" the training data. They took the simulated recordings and artificially added static, changed the speed, and stretched the sound waves.
  • Analogy: Imagine teaching a dog to recognize a "sit" command. If you only teach it in a quiet room, it might get confused in a windy park. So, you practice in the wind, with loud music, and while the dog is running. By the time you get to the park, the dog knows exactly what to do. The A-CNN was trained to be "deaf" to the static and "smart" enough to find the signal even when the data looks weird.

The Results: A 40% Boost

The results were impressive. By using this AI detective:

1. It kept 90% of the good whispers: It didn't throw away the real signals.
2. It silenced 50% of the shouting: It successfully identified and removed half of the background noise that the old methods couldn't catch.

Because the "noise" is cut in half, the experiment becomes 40% more sensitive.

• Analogy: Imagine you are trying to find a needle in a haystack. If you can magically remove half the hay, you don't just find the needle faster; you are 40% more likely to find it before you give up.

Why This Matters for the Future

Currently, building better detectors requires building bigger, cleaner, and more expensive tanks (hardware). This is like trying to hear the whisper by building a soundproof room, which costs millions of dollars.

This paper shows that we can get a 40% upgrade in performance just by changing the software. It's like upgrading your car's engine with a software update instead of buying a new car.

This technique will be crucial for the next generation of experiments (like XLZD), which will be massive. Without this AI trick, the background noise might drown out the signal entirely. With the A-CNN, these future detectors have a much better chance of finally hearing that cosmic whisper and solving one of the universe's biggest mysteries.

Summary

• The Goal: Find a rare particle decay to prove neutrinos are their own antiparticles.
• The Problem: The signal is drowned out by radioactive noise.
• The Solution: An AI model (A-CNN) that analyzes the detailed "sound" of particle interactions.
• The Secret Sauce: "Augmentation" (training the AI on messy, noisy data) so it works perfectly in the real world.
• The Payoff: A 40% increase in sensitivity without spending a dime on new hardware.

Technical Summary

1. Problem Statement

Context:
Dual-phase Liquid Xenon (LXe) Time Projection Chambers (TPCs), such as XENONnT, are primarily designed for dark matter (WIMP) detection. However, they also possess the potential to search for Neutrinoless Double-Beta Decay ($0\nu\beta\beta$) of $^{136}$Xe. Discovering this process would prove that neutrinos are Majorana particles and explain the matter-antimatter asymmetry in the universe.

The Challenge:

• Energy Scale Mismatch: Dark matter searches focus on ultra-low energies (<100 keV), whereas the $0\nu\beta\beta$ signal occurs at a high Q-value of 2.458 MeV.
• Background Dominance: At this high energy, the detector is plagued by material-induced backgrounds (gamma rays from radioactive isotopes like $^{214}$Bi in detector materials). Specifically, a gamma peak from $^{214}$Bi at 2447 keV sits directly on top of the expected $0\nu\beta\beta$ peak.
• Limitations of Traditional Methods: While hardware upgrades (ultra-pure materials, active vetoes) can reduce backgrounds, they are costly and difficult to implement on existing large-scale detectors. Traditional analysis cuts (e.g., Single-Site vs. Multi-Site event selection) are insufficient to fully reject these backgrounds without sacrificing significant signal efficiency.
• Goal: Develop a cost-effective software solution to improve background rejection and signal acceptance without additional hardware investment.

2. Methodology

The authors propose an Augmented Convolutional Neural Network (A-CNN) to classify events based on their topological signatures in the detector.

A. Data Source & Preprocessing:

• Detector: XENONnT (5.9 t sensitive LXe mass).
• Signal: $0\nu\beta\beta$ events are primarily Single-Site (SS) energy depositions.
• Background: Material gamma rays are primarily Multi-Site (MS) events (multiple scatterings).
• Input Data: The model uses S2 signals (ionization electroluminescence) detected by the top PMT array.
  • Instead of raw per-channel waveforms, the data is processed into 1D time series.
  • For events with multiple S2 peaks (main and alternative), the waveforms are resampled to a constant physical time step ($dt = 200$ ns), aligned by time difference, and concatenated into a single 1500-sample vector.
  • Waveforms are normalized to unit area.

B. Model Architecture (A-CNN):

• Type: A Convolutional Neural Network (CNN) designed for 1D time-series data.
• Structure:
  • Feature Extractor: Six convolutional blocks using LeakyReLU activation, Batch Normalization, and Max Pooling. This progressively reduces the time dimension while expanding channel depth (from 1 to 256 channels), resulting in a tensor of shape $[9, 256]$.
  • Classifier: A fully connected (linear) classifier that flattens the feature tensor and outputs a probability score (0 to 1) via a Sigmoid function, indicating the likelihood of an event being signal-like (SS).
• Training Data: Full-chain Monte Carlo (MC) simulations using GEANT4 and the XENONnT waveform simulator (WFSim).
  • 40,000 electron events (SS signal) and 40,000 gamma events (MS background).
  • Energies uniformly distributed between 0–4 MeV.

C. Data Augmentation Strategy:
To ensure the model trained on MC generalizes to real data, two specific augmentation techniques were applied:

1. Random Upsampling: Randomly stretching the width of the S2 waveform to make the model robust against variations in S2 width between simulation and reality.
2. Random Noise Injection: Adding Gaussian noise to the time-series input to prevent the model from relying on baseline suppression artifacts caused by energy normalization, thereby mitigating energy bias.

3. Key Contributions

1. Novel Algorithm: Introduction of an A-CNN specifically tailored for LXe TPC event topology classification using low-level S2 waveform data.
2. Robustness via Augmentation: Demonstration that targeted data augmentation (resampling and noise injection) effectively bridges the gap between MC simulations and real detector data, reducing Data-MC discrepancies to ~3% (superior to previous experiments like EXO-200 which had ~10% discrepancies).
3. Interpretability: Use of Saliency Maps to visualize the model's decision-making process. The study confirms the model focuses on physically relevant features (rising/falling edges of the S2 peak) rather than random noise, validating that the discrimination is based on genuine event topology.
4. Cost-Effective Sensitivity Boost: Proving that software-based analysis can significantly enhance physics reach without requiring expensive hardware upgrades.

4. Results

Classification Performance:

• ROC Curve: The model achieved an Area Under the Curve (AUC) of 0.85.
• Operating Point: At 90% Signal Acceptance, the model achieves 62% Background Rejection on the test set.
• Post-Traditional Cuts: Even after applying standard data quality and S2 SS cuts, the A-CNN still rejects an additional 50% of the remaining background in the $Q_{\beta\beta}$ region while maintaining 90% signal efficiency.

Data-MC Validation:

• The model was validated on three real-world datasets: $^{136}$Xe $2\nu\beta\beta$ (signal-like), $^{208}$Tl gamma rays (background-like), and $^{212}$Pb $\beta+\gamma$ events.
• The A-CNN score distributions between real data and MC showed excellent agreement (residual differences <3%), confirming the model's reliability for physics analysis.

Sensitivity Improvement:

• Counting Experiment: The background rejection translates to a 27.3% improvement in sensitivity ($S/\sqrt{B}$).
• Likelihood Fit (Toy MC): A more rigorous likelihood fit simulation projected a 39.4% (approx. 40%) improvement in the projected median lower limit for the $^{136}$Xe $0\nu\beta\beta$ half-life ($T_{1/2}^{0\nu\beta\beta}$).
• Comparison: This brings the projected sensitivity of XENONnT significantly closer to dedicated $0\nu\beta\beta$ experiments like KamLAND-Zen, despite XENONnT's primary design for dark matter.

5. Significance and Outlook

• Immediate Impact: The A-CNN method allows XENONnT to maximize its physics potential for $0\nu\beta\beta$ searches using existing data and infrastructure.
• Future Applications: The technique is directly applicable to next-generation multi-tonne detectors like XLZD (60–80 t target mass). It offers a scalable, cost-effective path to achieving world-leading sensitivity for $0\nu\beta\beta$ without compromising the primary WIMP search mission.
• Broader Implication: This work demonstrates that deep learning, when combined with rigorous data augmentation and interpretability checks, can solve complex background rejection problems in rare-event physics, potentially replacing or augmenting traditional hardware-centric background suppression strategies.
• Future Work: The authors plan to explore spatiotemporal deep learning models that utilize full per-channel timing and topology information to further enhance discrimination power.