Vision Calorimeter for High-Energy Particle Detection

The paper proposes Vision Calorimeter (ViC), a novel framework that adapts visual object detectors with a physics-inspired heat-conduction operator to significantly improve the accuracy of anti-neutron position and momentum estimation in high-energy physics.

The Gist

Imagine a giant, high-tech camera that doesn't take pictures of people or landscapes, but instead captures the invisible "shadows" left behind when tiny particles smash into each other at nearly the speed of light. This is the job of a device called an Electromagnetic Calorimeter (EMC).

The problem is that these "particle shadows" look nothing like normal photos. Instead of clear shapes, they look like a sparse, scattered constellation of dots on a dark background. Trying to figure out exactly where a specific particle hit and how fast it was moving just by looking at these scattered dots is like trying to guess the location and speed of a firework just by looking at a few stray sparks in a dark field.

The researchers in this paper, led by Hongtian Yu, decided to solve this by borrowing a trick from the world of self-driving cars and security cameras.

The Big Idea: Teaching a "Traffic Cop" to See Particles

In computer vision (the field that lets computers "see"), there are smart programs called Object Detectors. These are usually trained to spot cars, dogs, or people in photos. They are very good at finding where an object is and what it is.

The team asked: What if we taught one of these "traffic cop" programs to spot anti-neutrons (a type of particle) in these weird particle images?

They created a system called Vision Calorimeter (ViC). Think of ViC as a translator that turns the messy, scattered "particle sparks" into a format that a standard computer vision brain can understand.

The Secret Sauce: The "Heat" Operator

The main challenge is that particle images are "discrete" (scattered dots) while normal photos are "continuous" (smooth gradients). To bridge this gap, the team invented a special tool called the Heat-Conduction Operator (HCO).

Here is the analogy:

• Normal Photos: Imagine a smooth, warm blanket. The heat is spread out evenly.
• Particle Images: Imagine a blanket with only a few hot spots and mostly cold areas.

The HCO acts like a magical heat diffuser. It takes those scattered "hot spots" (the particle energy) and simulates how heat would naturally spread out through a material. By doing this mathematically (using a technique called Discrete Cosine Transform), it turns the scattered dots into a smooth, continuous pattern that looks much more like a normal photo.

This allows the computer to use its pre-existing "knowledge" of how to see shapes, even though it's looking at particle data for the first time.

How It Works in Practice

1. The Setup: They took data from the BESIII experiment (a real particle collider). They mapped the energy readings from the detector cells onto a 2D grid, creating a "particle image."
2. The Training: They taught the ViC system to act like a detective. Instead of just saying "there is a particle here," it had to answer two questions:
   • Where did it hit? (Position)
   • How fast was it going? (Momentum)
3. The Innovation: Since they didn't have perfect "bounding boxes" (rectangles drawn around the particles) to teach the AI, they invented a way to create "fake" but accurate boxes based on the physics of how the energy spreads.

The Results: A Huge Leap Forward

The paper claims that ViC is a massive improvement over the old ways of doing this:

• Better Positioning: The old methods (called "clustering algorithms") were like guessing the location of a firework with a 17-degree error. ViC reduced this error to just 9 degrees. That's a 46% improvement in accuracy.
• Speed Detection (The First Time): Perhaps most importantly, this is the first time a method has successfully estimated the momentum (speed) of these anti-neutrons using only this type of detector. The error rate for speed was about 21%, which is a significant breakthrough.
• Real-World Proof: They tested the system by reconstructing a known particle event (a J/ψ particle decaying). The system successfully recreated the "fingerprint" of this event, proving it works for real physics analysis.

In a Nutshell

The researchers took a problem that was too messy for traditional math to solve, turned the data into a picture, and used a "heat-diffusing" filter to make that picture look like something a standard AI could understand. The result is a system that can pinpoint where particles hit and how fast they were moving with much greater accuracy than ever before, acting as a powerful new tool for physicists to understand the fundamental building blocks of the universe.

Technical Summary

Technical Summary: Vision Calorimeter (ViC) for High-Energy Particle Detection

1. Problem Statement

In high-energy physics, accurately estimating the parameters (incident position and momentum) of anti-neutrons ($\bar{n}$) is a critical yet challenging task. Unlike charged particles, anti-neutrons are electrically neutral and do not leave tracks in tracking detectors; they can only be detected via their interaction with the Electromagnetic Calorimeter (EMC). However, the energy deposition patterns of $\bar{n}$ in the EMC are characterized by discreteness and sparsity, with scattered regions of foreground activation.

Conventional methods, primarily based on analytical clustering algorithms (e.g., ClustAlgo), struggle to reliably distinguish $\bar{n}$ from background noise and other particles. These methods often fail to accurately determine the incident position and cannot effectively regress the incident momentum. While deep learning models have shown promise in particle physics, standard visual object detectors are not naturally suited for these sparse, non-natural image patterns, creating a significant domain gap between natural images and high-energy particle images.

2. Methodology: Vision Calorimeter (ViC)

The authors propose Vision Calorimeter (ViC), an end-to-end deep learning framework that adapts visual object detection architectures to analyze particle images derived from EMC data. The methodology consists of three core components:

A. Data Formatting and Representation

The authors establish a mapping between the cylindrical arrangement of EMC cells and a 2-D image plane.

• Image Construction: The spherical coordinates ($\phi, \theta$) of the EMC cells are projected onto a 2-D image (960 $\times$ 480 pixels).
• Energy-to-Color Mapping: To handle the highly imbalanced energy distribution (where most values are low-energy), the authors apply a logarithmic transformation followed by a specific mapping to RGB channels. Low, medium, and high energy ranges are mapped to Blue, Green, and Red channels, respectively, with histogram equalization applied to mitigate extreme values.

B. The Heat-Conduction Operator (HCO)

To bridge the distribution gap between natural and particle images, ViC introduces a physics-inspired Heat-Conduction Operator (HCO) into the backbone and head of the detector.

• Mechanism: Inspired by the physical process of heat diffusion, the HCO models feature extraction using the 2-D Discrete Cosine Transform (DCT) and its inverse (IDCT). It maps complex features to "high-temperature" regions (accumulation) and sparse features to "low-temperature" regions (dissipation).
• Frequency Alignment: By transforming spatial domain signals (discrete pulses) into the frequency domain, the HCO makes particle image features more akin to natural image features, facilitating the transfer of pre-trained visual representations.
• HCO-K (Improved): The authors propose an enhanced version, HCO-K, where the heat conductivity coefficient ($k$) is made dependent on the input feature map via feature fusion. This allows the model to dynamically capture distinct conduction attributes across different samples and layers, rather than relying on a shared, static embedding. The backbone network, named vHeatK, stacks these HCO-K layers following a hierarchical design similar to the Swin Transformer.

C. Detection Strategy and Head Design

• Pseudo Bounding Box Generation: Since precise ground-truth bounding boxes are unavailable for particle images, the authors propose a strategy to generate pseudo bounding boxes. The incident position is treated as the center, and the box size is defined as a multiple of the cell size (optimized at $10\times$ cell size to cover $\approx 99\%$ of energy deposition).
• Dual-Attention Architecture: Recognizing the conflicting requirements for position and momentum, ViC employs a specialized head structure:
  • Position Prediction: Utilizes local attention to focus on the specific cluster of activated cells near the incident point.
  • Momentum Regression: Utilizes global attention to capture the entire energy distribution, as every deposition point potentially carries momentum information.
• Loss Functions: The framework uses standard object detection losses for localization and a specific regression loss for momentum, ensuring the predicted momentum remains positive via an exponential output layer.

3. Key Contributions

1. First End-to-End Baseline: ViC is presented as the first end-to-end deep learning baseline for anti-neutron detection using solely EMC data, successfully migrating visual object detectors to high-energy particle images.
2. Physics-Inspired Representation: The introduction of the Heat-Conduction Operator (HCO) and its improved variant HCO-K specifically tailored to the discrete and radial patterns of particle incidence. This module effectively bridges the domain gap between natural and particle images.
3. Dual-Task Optimization: The framework uniquely addresses the trade-off between local and global information needs by combining local attention for position prediction and global attention for momentum regression within a unified detection pipeline.
4. Pseudo-Annotation Strategy: A novel method for generating pseudo bounding boxes from point-based ground truths, enabling the use of standard object detection training pipelines.

4. Experimental Results

Experiments were conducted on the $\bar{n}$-1m dataset, containing 986,343 electron-positron collision events from the BESIII experiment.

• Incident Position Prediction: ViC achieved a Mean Angular Bias (mAB) of 9.32°, representing a 46.16% reduction in error compared to the conventional ClustAlgo method (17.31°). It also outperformed other deep learning baselines, including SFNet (12.87°), Set Transformer, and standard object detectors like RetinaNet and DINO.
• Incident Momentum Regression: ViC provided the first baseline result for momentum regression in this context, achieving a Mean Relative Error (mRE) of 21.48%. This performance surpasses other calorimeters in sub-GeV energy regions, where hadronic calorimeters typically exhibit errors exceeding 50%.
• Physical Analysis Validation: The authors demonstrated the utility of ViC in reconstructing the invariant mass of $J/\psi$ particles decaying into a proton, pion, and anti-neutron. The results showed a sharp peak in the invariant mass spectrum correctly centered at the nominal $J/\psi$ mass, validating the method's capability for downstream physics analysis.
• Ablation Studies: Results confirmed that the vHeatK backbone outperforms ResNet, ConvNeXt, and Swin Transformers. The HCO-K operator and the specific head architecture (combining local and global attention) were shown to be critical for performance.

5. Significance and Claims

The paper claims that ViC demonstrates the great potential of adapting advanced machine learning technologies to natural science domains, specifically high-energy physics.

• Reliability: ViC is presented as a reliable particle detector that significantly outperforms conventional analytical methods in parameter estimation.
• Novel Capability: The study highlights a breakthrough in enabling the measurement of incident momentum for anti-neutrons using only EMC data, a task previously considered difficult or impossible with standard algorithms.
• Scalability: The framework suggests a pathway for modeling the extensive data generated by high-energy colliders.
• Limitations and Scope: The authors modestly note that the current validation is specific to the BESIII experiment conditions (momenta below 1.2 GeV/c and absence of pile-up). They acknowledge that extending this approach to other high-energy physics experiments with different data access policies and environmental conditions remains a direction for future work.

The code is made available to the community to facilitate further research in this domain.