What Machine Learning Can Do for Focusing Aerogel Detectors

This paper presents machine learning-inspired approaches, adapted from computer vision, to filter ambient background hits in the Focusing Aerogel Ring Imaging CHerenkov (FARICH) detector at the Super Charm-Tau factory, thereby reducing data flow and improving particle velocity resolution.

The Gist

Imagine you are trying to take a clear photograph of a single firefly blinking in a dark field. Now, imagine that instead of a quiet night, you are standing in the middle of a massive, chaotic fireworks display. Every time you try to snap a picture, thousands of random sparks (noise) light up the camera sensor, making it nearly impossible to see the one firefly you care about (the signal).

This is the exact problem facing scientists working on a new particle detector called FARICH, which is being built for a massive physics experiment called the Super Charm-Tau factory. Their goal is to identify specific subatomic particles by looking at the faint rings of light they create. However, because of where the detector is located, it gets bombarded with so much "background noise" (random hits) that the real signal gets drowned out. The ratio of noise to signal is roughly 70 to 1.

Here is how the authors used Machine Learning (ML) to solve this, explained simply:

1. The Old Way vs. The New Way

The Old Way (The Rulebook):
Traditionally, scientists tried to filter out noise by writing strict mathematical rules based on physics. For example, they might say, "If a hit happens at exactly 3 nanoseconds, keep it; if it's at 4, throw it away."

• The Problem: This is like trying to sort a messy room by only looking at the color of the objects. It works okay if the room is only slightly messy, but if the room is overflowing with junk (heavy background noise), these rigid rules fail. They also struggle to adapt if you add new types of data.

The New Way (The Smart Eye):
The authors decided to use Machine Learning, specifically techniques borrowed from computer vision (the technology that lets computers "see" and recognize objects in photos).

• The Analogy: Instead of following a rulebook, they trained a computer to "look" at the data like a human looks at a crowded photo. The computer learns to recognize the shape and pattern of the real signal, ignoring the random chaos around it, just as you can spot a friend in a crowd even if they are wearing a different hat than usual.

2. How They Taught the Computer

To train this "smart eye," the researchers created a digital simulation (a video game version of the detector) using a tool called Geant4.

• The Input: They fed the computer a special "image" made of two layers:
  1. Where the light hits (coordinates).
  2. When the light hits (time).
• The Pattern: Real signals tend to cluster together tightly in time (like a group of friends huddled together), while the noise is scattered randomly (like people walking alone in different directions).
• The Training: They showed the computer millions of these "images," some with the real signal and some with just noise. The computer (using a specific type of neural network called ResNet-18) learned to distinguish the "huddled friends" from the "random walkers."

3. The Results: A Cleaner View

The results were impressive. When they tested the system with a high level of noise (simulating the worst-case scenario):

• Noise Reduction: The system successfully filtered out 90% of the background noise.
• Signal Retention: It kept 95% of the real, important signals.

Think of it as a bouncer at a club who is so good at spotting the VIPs that they let 95% of the VIPs in while kicking out 90% of the people just trying to crash the party.

4. Where It Works Best (and Where It Struggles)

The "smart eye" works best when the particles are moving fast (high momentum). However, just like a human might struggle to see a firefly if it's moving very slowly or at a weird angle, the system's performance drops slightly when particles are slow or hit the detector from a sharp angle.

5. The Big Picture

The paper concludes that while traditional math rules are good for simple situations, Machine Learning is a powerful tool for messy, noisy environments. By treating the detector data like an image and using computer vision techniques, they can clean up the data much more effectively. This not only helps the current experiment but could also be used for other detectors in the future, like the one planned for the NICA facility.

In short: They replaced a rigid rulebook with a "smart camera" that learned to ignore the fireworks so it could finally see the firefly.

Technical Summary

Technical Summary: Machine Learning for Focusing Aerogel Detectors

Problem Statement
The Super Charm-Tau factory (SCTF) experiment requires reliable particle identification (PID), for which a Focusing Aerogel Ring Imaging CHerenkov (FARICH) detector is proposed. A critical challenge for the FARICH is its operating environment, where cooling constraints lead to a high rate of ambient background hits. Specifically, Silicon Photomultiplier (SiPM) properties and operating temperatures can result in background hit rates of approximately $1 \text{ MHz/mm}^2$, yielding a signal-to-background ratio as low as 0.014.

Conventional pattern recognition methods, which rely on empirical statistics derived from physical properties (e.g., hit times, cell occupancy), are effective only under moderate background conditions. These statistical approaches struggle in the heavy background environment of the FARICH due to their local nature and the requirement for prior particle track information. Furthermore, integrating new data sources into these statistical frameworks is difficult. The primary objective is to mitigate this noise to reduce data flow and improve particle velocity resolution without relying on pre-existing track information.

Methodology
The authors propose a Machine Learning (ML) approach inspired by computer vision techniques to filter signal hits and reconstruct events.

• Data Representation: Signal data is generated via Geant4 simulation, containing photon coordinates ($x_c, y_c$) and hit times ($t_c$) on a SiPM grid. To leverage ML capabilities for unified data handling, the input is formatted as a 2-channel image:
  1. A bilinearly interpolated binary mask derived from coordinates.
  2. Normalized hit times for corresponding pixels.
     This format exploits the temporal compactness of signal hits (clustered within $\sim 2$ ns) compared to the full event duration ($\sim 7$ ns). Uniform random noise is superimposed to simulate ambient background.
• Model Architecture: A ResNet-18 convolutional neural network (CNN) is employed, with modifications to the input and output layers to accommodate the specific data format. The model is initialized with pretrained weights from a reconstruction task to accelerate learning.
• Task Formulation: The noise filtering problem is framed as a binary classification task (signal present vs. signal absent).
  • Ground Truth: Events are labeled positive if they contain $\ge 10$ signal photons within a bounding box derived from particle track parameters.
  • Training Strategy: The authors investigate weighted binary classification (Eq. 1) to control the trade-off between efficiency (True Positive Rate) and noise reduction (False Positive Rate). They compare training with specific positive class weights ($w_{pos}$) against simply adjusting the decision threshold in a standard logistic regression setup.
  • Alternative Approach: The paper also considers bounding box regression (predicting coordinates of the signal box) as an alternative to classification.

Key Results

• Noise Rejection Efficiency: The ML-based approach achieves significant noise rejection while maintaining high signal efficiency. For $\pi^-$ events at a noise frequency of $10^6 \text{ Hz/mm}^2$, the model achieves 90% background rejection while retaining 95% of the signal.
• Training Strategy Comparison: No significant performance difference was observed between training with specific positive class weights and using a standard classifier with a tuned threshold. Consequently, the authors conclude that the simpler method (threshold tuning) is preferable to avoid the computational cost of training multiple models.
• Performance Dependencies:
  • The model performs best for high-momentum events.
  • Performance degrades at the lower end of the momentum spectrum and for angles of incidence closer to the tangent.
  • The classification approach is found to be less precise but significantly more robust than bounding box regression. However, chaining bounding box regression with classification for positive events could potentially yield even higher noise reduction.
• Metrics: Performance is evaluated using Receiver Operating Characteristic (ROC) curves and the Area Under Curve (AUC), with a focus on efficiency and noise event reduction rates.

Significance and Claims
The paper claims that applying Machine Learning to the FARICH detector addresses the limitations of conventional statistical methods in high-background environments. By extracting high-level features from unified data sources (coordinates and time), ML offers a robust solution for online noise filtering and offline reconstruction.

The authors highlight that while the initial results are promising, the research is part of a broader effort to solve essential problems in the FARICH development, including fast online filtering, offline reconstruction, and fast simulation. They note that the applicability of this research extends beyond the SCTF, as the FARICH design is also under discussion for the Spin Physics Detector (SPD) at NICA. The work suggests that ML has great potential for RICH detectors in high-energy physics, offering a pathway to handle the specific noise challenges posed by SiPM-based aerogel detectors.