Readout and PID using AIML for SoLID High Background Cherenkov Detectors

This paper presents the development of a high-rate MAROC sum readout system and artificial intelligence-based particle identification methods that significantly improve $\pi/K$ separation for the SoLID Cherenkov detectors at Jefferson Lab compared to traditional methods.

The Gist

Imagine you are at a massive, crowded music festival. Thousands of people are dancing, shouting, and moving all at once. In the middle of this chaos, you are trying to spot one specific person wearing a bright neon hat.

This paper describes how scientists at Jefferson Lab are building a "super-eye" (a detector) to do exactly that, but instead of people, they are looking for tiny subatomic particles flying through a machine at nearly the speed of light.

Here is the breakdown of how they solved this problem using two main "superpowers": Better Vision and Artificial Intelligence.

1. The Problem: The "Noisy Party"

The scientists are studying a device called SoLID. It’s designed to smash particles together to understand the "glue" that holds atoms together.

The problem is that the environment is incredibly "noisy." Imagine trying to count how many people are wearing neon hats while a thousand other people are throwing confetti and flashing strobe lights everywhere. In the detector, "confetti" is actually background radiation that looks a lot like the particles the scientists actually want to study. If they just count the total amount of light, they can’t tell the difference between a "real" particle (a Pion) and "background noise" (a Kaon or an electron).

2. The Solution Part 1: The "High-Definition Camera" (The Readout)

Previously, detectors were like old cameras that just took one blurry, bright photo of the whole scene. You could see that something happened, but not where.

The researchers built a new electronic system called the MAROC sum readout.

• The Old Way: A single light sensor that says, "I saw a flash of light!"
• The New Way: A high-speed, high-definition sensor that says, "I saw a flash of light, and it hit exactly these 64 tiny spots in a specific pattern."

By breaking the light down into pixels (tiny dots), quadrants (groups of dots), and total sums, they aren't just seeing a flash; they are seeing a shape. It’s the difference between hearing a loud bang in a dark room and seeing the specific shape of a firework exploding in the sky.

3. The Solution Part 2: The "Smart Security Guard" (AIML)

Even with a high-def camera, the "confetti" (background noise) is still flying everywhere. If you just look for "brightness," you'll get fooled.

To solve this, they used Artificial Intelligence (AIML). They trained a digital "brain" (a neural network) using computer simulations.

• They showed the brain millions of examples: "This pattern of dots is a Pion (the good guy)," and "This messy, scattered pattern is just background noise (the bad guy)."
• Because the AI can look at the spatial pattern (the shape of the light ring) rather than just the brightness, it can tell the difference.

The Analogy:
Imagine you are looking at a rainy window at night. A simple sensor would just say, "It's dark and wet." But the AI is like a trained eye that can look at the streaks of water and say, "Those streaks are just rain, but that specific circular pattern in the middle is actually a car's headlight."

The Result: A Success!

The scientists tested this "Smart Eye" and found that:

1. The Hardware works: The electronics can handle the "high-speed party" without crashing or getting overwhelmed.
2. The AI is brilliant: Using the new pixel-based information, the AI was able to correctly identify the particles more than 90% of the time, even when the "confetti" was flying thick.

In short: By upgrading from a "blurry light sensor" to a "high-speed pattern recognizer" and adding an "AI brain" to interpret the shapes, they have found a way to see clearly through the chaos of the subatomic world.

Technical Summary

Technical Summary: Readout and PID using AIML for SoLID High Background Cherenkov Detectors

1. Problem Statement

The Solenoidal Large Intensity Device (SoLID) at Jefferson Lab is designed for high-luminosity experiments to study nucleon structure. A critical challenge lies in the Heavy Gas Cherenkov (HGC) detector, which must perform $\pi/K$ (pion/kaon) separation in an extremely high-rate, high-background environment.

The primary difficulties are:

• High Background: Low-energy electrons (from beam interactions and knock-on effects) produce Cherenkov photons that mimic signal events.
• Rate Requirements: The detector must handle rates up to 200 kHz per pixel or 4 MHz per Photomultiplier Tube (PMT).
• PID Limitations: Traditional particle identification (PID) methods, such as simple photoelectron-counting (NPE) cuts, fail to distinguish between pions and kaons when background levels are high, as the signal and background distributions overlap significantly.

2. Methodology

The researchers addressed these challenges through two parallel developments: hardware (readout electronics) and software (AI-based pattern recognition).

A. Hardware: MAROC Sum Readout System
To capture the spatial information of Cherenkov rings (which helps distinguish signal from random background), the team designed a custom readout system using Multi-Anode Photomultiplier Tubes (MAPMTs).

• Architecture: They developed a new "sum board" integrated with MAROC ASICs. This allows for the simultaneous readout of three distinct signal levels:
  1. Pixel-level: Individual hits via Time-to-Digital Converters (TDC).
  2. Quadrant-sum: Analog sums of 16-pixel groups.
  3. Total-sum: The sum of all 64 pixels in a PMT.
• Testing: The system underwent saturation studies, TDC threshold optimization, and high-rate bench tests (reaching ~6 MHz/PMT), as well as "Ring Tests" using LEDs and cosmic muons to verify signal linearity and ring reconstruction accuracy.

B. Software: AIML-based Particle Identification
To leverage the spatial data provided by the new electronics, the team implemented Artificial Intelligence and Machine Learning (AIML).

• Model: They utilized a Multilayer Perceptron (MLP) neural network consisting of an input layer, five hidden dense layers (with ReLU activation and 15% dropout for generalization), and a sigmoid output layer.
• Training: The models were trained on realistic Geant4 simulations that included beam-related backgrounds. They trained separate models for PMT-only, quad-sum, and pixel-level data to compare the efficacy of different readout granularities.

3. Key Contributions

• Novel Readout Design: The creation of a MAROC sum readout system that preserves high-granularity spatial information while maintaining the ability to provide high-speed analog sums for triggering.
• High-Rate Validation: Demonstration that the custom electronics can sustain rates exceeding the expected SoLID operational requirements (up to 6 MHz/PMT).
• Advanced PID Framework: Transitioning from traditional 1D photoelectron counting to a multi-dimensional pattern recognition approach using supervised learning.

4. Results

• Electronics Performance: The MAROC sum system showed excellent linearity and successfully reconstructed Cherenkov rings using the Hough transform algorithm. The agreement between pixel-scaler rates and sum-scaler rates was within 3%.
• PID Efficiency:
  • Traditional Method: A simple NPE cut achieved only ~50–60% efficiency for both pions and kaons in high-background scenarios.
  • AIML Method: The MLP models significantly outperformed the NPE cut.
  • Granularity Impact: While PMT-only readout barely met the target (90.82% kaon efficiency in some cases), the quad and pixel readout schemes achieved pion and kaon efficiencies above 90% across various kinematic settings.
  • Robustness: The models showed high stability (uncertainties < 1%) and high Area Under the Curve (AUC) values, indicating strong classification power.

5. Significance

This work provides a validated technical roadmap for the SoLID experiment. By combining high-granularity, high-rate electronics with deep-learning-based pattern recognition, the researchers have demonstrated that it is possible to achieve robust particle identification even in the presence of intense beam-related backgrounds. This approach moves Cherenkov detector technology from simple "photon counting" to sophisticated "pattern recognition," which is essential for the next generation of high-luminosity nuclear physics experiments.