Data-Driven Calibration of Large Liquid Detectors with Unsupervised Learning

This paper presents a novel unsupervised deep learning method that utilizes a simplified physical model of optical photon transport to automatically extract photomultiplier tube calibration constants from radioactive background events in the large-scale SNO+ liquid scintillation detector.

The Gist

Imagine you are trying to take a group photo of 9,300 people (the Photomultiplier Tubes, or PMTs) inside a giant, dark swimming pool (the liquid scintillator detector). When a particle zips through the water, it creates a tiny flash of light, like a firework. To figure out exactly where that firework happened, you need to know exactly when each person saw it.

The problem? Everyone's watch is slightly off. Some people are slow to react, some have thick glasses that delay the light, and some have different cables connecting them to the camera. If you don't fix these timing differences, your photo will be blurry, and you won't know where the firework actually was.

Traditionally, scientists fix this by sending in a "calibration crew" with a special laser ball. They turn on the laser at known spots and times, and the scientists manually adjust everyone's watches. But this is expensive, takes a lot of time, and sometimes you can't do it because the detector is busy doing real science.

This paper introduces a clever new way to fix the watches without ever sending in a human or a laser.

Here is how they did it, using a few simple analogies:

1. The "Ghost" in the Machine

Instead of using a laser, the scientists used the "ghosts" that are already haunting the detector: natural radioactive background noise. Specifically, they looked at tiny, natural energy bursts (from Polonium-210) that happen all the time inside the tank.

Think of these as tiny, random fireflies appearing in the dark pool. We don't know exactly where they are or exactly when they blinked, but we know they are there.

2. The "Blind" Detective (Unsupervised Learning)

Usually, to teach a computer to solve a puzzle, you give it the answer key (like "This firefly was at position X at time Y"). But here, the scientists didn't have the answer key. They used Unsupervised Learning.

Imagine a detective trying to solve a crime by looking at a chaotic scene. The detective doesn't know who did it or where, but they know the rules of physics:

• Light travels at a constant speed.
• If the firefly was in the center, everyone should see it at roughly the same time.
• If the firefly was on the left, the people on the left should see it first.

The computer (a Deep Learning model) acts like this detective. It looks at the chaotic data from the 9,300 PMTs and asks: "If I tweak everyone's watch settings just a tiny bit, can I make all these fireflies look like they happened in a straight line?"

3. The "Time Walk" Problem

There's a tricky part called "Time Walk." Imagine a person with a slow reaction time who only notices a firefly if it's really bright. If the firefly is dim, they wait longer to react. If it's bright, they react instantly.

The computer had to learn a specific rule for every single PMT: "If the signal is weak, add 2 nanoseconds to the time. If it's strong, add 0." It had to figure out this rule for over 7,500 different tubes simultaneously.

4. The "Symphony" Tuning

The scientists set up a massive feedback loop.

1. The computer guesses the timing settings for all the PMTs.
2. It uses those settings to guess where the "fireflies" (radioactive events) are.
3. It checks if the timing makes sense. If the firefly looks like it exploded in two places at once, the computer knows, "Oops, my timing guesses are wrong."
4. It adjusts the timing settings slightly and tries again.

They did this millions of times. It's like tuning a massive orchestra of 9,300 instruments. Instead of a conductor telling each musician when to play, the computer listens to the whole group and says, "You're a little sharp, you're a little flat," until the whole group plays in perfect harmony.

The Result

By the end of the process, the computer had figured out the exact timing quirks for every single PMT, using only the natural background noise.

• It was faster: It took about a day on a powerful computer chip, compared to weeks of manual laser work.
• It was accurate: It was just as good as, and in some ways better than, the old laser method.
• It was a "Magic Eye" test: Because the computer didn't know the answers beforehand, it actually found a hidden problem in the detector's electronics that the old laser method missed!

In short: The scientists taught a computer to tune a giant, complex machine by letting it listen to the background noise and figure out the rhythm on its own. It's a brilliant example of using "smart" software to do the heavy lifting that used to require expensive hardware and human labor.

Technical Summary

Here is a detailed technical summary of the paper "Data-Driven Calibration of Large Liquid Detectors with Unsupervised Learning."

1. Problem Statement

Large liquid scintillator detectors, such as SNO+, utilize thousands of Photomultiplier Tubes (PMTs) to detect light from charged particles. Accurate event reconstruction relies heavily on precise PMT timing calibration (typically to within a few tenths of a nanosecond). Two primary challenges exist:

• Time Walk Effect: The time at which a PMT triggers depends on the charge of the signal. Higher charge pulses cross the threshold earlier than lower charge pulses, creating a charge-dependent time delay.
• Limitations of Traditional Calibration: Standard methods rely on dedicated hardware campaigns (e.g., laserballs or diffuser balls) deployed inside the detector. These campaigns are infrequent, labor-intensive, require stopping physics data-taking, and often use light sources with different spectral/timing characteristics than actual physics events, necessitating extrapolation.

The paper proposes a method to extract calibration constants directly from physics data (specifically radioactive background events) using unsupervised deep learning, eliminating the need for dedicated calibration hardware campaigns.

2. Methodology

The core approach frames the calibration problem as a large-scale regression task solved via unsupervised deep learning.

A. Physical Model

The authors model the PMT timing behavior using three parameters per PMT ($i$):

1. Constant Delay ($c_i$): Electronic and cable delays.
2. Time Walk Parameters ($a_i, b_i$): A decaying exponential function modeling the charge-dependent time walk:
   $$\tau_i(q) = a_i \exp(-q/b_i) + c_i$$
   where $q$ is the integrated PMT charge.
   • Constraint: To ensure numerical stability, $b_i$ is constrained to be positive using a "softplus" transformation.
   • Scale: With ~7,500 online PMTs, this results in over 22,000 free parameters to optimize.

B. Data Source

The method utilizes $^{210}\text{Po}$ alpha decays (Q = 5.4 MeV, visible energy ~0.45 MeV) from natural radon contamination in the SNO+ detector.

• Advantage: Alpha particles create point-like energy depositions in the scintillator, serving as ideal calibration sources.
• Dataset: Approximately 7 million events collected over 6 days, providing ~100,000 hits per PMT.

C. Unsupervised Learning Architecture

Since the true event vertex and time are unknown, a Neural Network (NN) is trained simultaneously with the calibration parameters.

• Architecture: A Transformer Encoder (6 layers, 4 attention heads) is used.
  • Input: PMT IDs and uncalibrated hit times.
  • Preprocessing: Hit times are centered by the event mean to ensure invariance to global time offsets. No sequence position encoding is used to maintain permutation invariance.
  • Output: Reconstructed event vertex ($x, y, z$) and time ($t$).
• Optimization Loop:
  1. Raw hit times are corrected by the current candidate calibration model ($\tau_i(q)$).
  2. The corrected times are fed into the Transformer to predict the event vertex and time.
  3. Time Residuals ($t_{res}$) are calculated by comparing predicted times against the time-of-flight (straight-line approximation) from the predicted vertex.
  4. Loss Function: The negative log-likelihood of the time residuals is calculated using a Jones and Faddy skew-t distribution (fitted to Monte Carlo data). This distribution models the sharp peak of prompt light and the long tail of scintillator re-emission.
  5. Backpropagation: Gradients flow through both the Transformer (to improve vertex reconstruction) and the calibration parameters (to minimize time residuals), optimizing them simultaneously.

D. Implementation

• Framework: PyTorch with automatic differentiation.
• Hardware: Single NVIDIA H100 GPU.
• Training Time: ~24 hours for 500,000 batches (batch size 2048).
• Optimizer: Adam with a 1-cycle learning rate scheduler.

3. Key Contributions

1. Novel Calibration Paradigm: Demonstrates the first successful application of unsupervised deep learning to extract PMT timing constants directly from physics data without dedicated calibration sources.
2. Scalability: Successfully optimizes >22,000 parameters simultaneously, a scale previously difficult to manage with traditional fitting methods.
3. End-to-End Training: The calibration procedure trains the position reconstruction network directly on data, removing the need for explicit Monte Carlo truth labels for the NN training phase.
4. Simplified Physics Modeling: Achieves high accuracy despite using simplified assumptions (e.g., straight-line time-of-flight ignoring refraction at the acrylic-water interface), proving the robustness of the data-driven approach.

4. Results and Validation

A. Monte Carlo (MC) Validation

• Accuracy: The method achieved a timing calibration accuracy of 0.14 ns FWHM (Full Width Half Maximum) when compared to ground truth in MC simulations.
• Stability: Residuals showed no significant dependence on PMT azimuthal angle or vertical position (except for known optical effects like the detector neck).
• Reproducibility: Results were consistent across different random seeds and optical parameter variations.

B. Real Data Validation

1. Electronic Structure Consistency: The derived delay constants ($c_i$) perfectly replicated the known electronic crate/card/channel structure of the detector, confirming the model learned physical realities rather than noise.
2. Comparison with Laserball: Compared to the standard SNO+ laserball calibration, the data-driven method showed a linear delay difference of ~1 ns from top to bottom. This was attributed to the ~10 cm uncertainty in the absolute position of the laserball, validating the data-driven method's precision.
3. Position Resolution ($^{214}\text{BiPo}$): Using $^{214}\text{Bi}$-$^{214}\text{Po}$ coincidences as a benchmark, the data-driven calibration improved the reconstructed distance ($\Delta R$) between decay vertices by a factor of 0.97 compared to the standard calibration, indicating a slight but measurable improvement in position resolution.
4. Anomaly Detection: The method successfully identified a previously overlooked timing deviation in "Crate 11" of the electronics, which was later confirmed by a lower-level electronics calibration. This highlights the method's utility for continuous detector monitoring.

5. Significance

This work represents a paradigm shift in the calibration of large-scale liquid scintillator detectors. By leveraging unsupervised deep learning, the authors have demonstrated that:

• Resource Efficiency: High-precision calibration can be achieved without interrupting physics data-taking or deploying complex hardware.
• Generalizability: The approach is applicable to any large detector with sufficient event statistics and can be extended to other applications beyond PMT timing.
• Continuous Monitoring: The speed and automation of the method allow for frequent re-calibration, enabling real-time diagnosis of detector health and electronic drifts.

The paper concludes that this data-driven, unsupervised approach provides a robust, accurate, and scalable alternative to traditional calibration campaigns, potentially becoming a standard tool for future neutrino and dark matter experiments.