Reconstruction of overlapping electromagnetic showers in calorimeters using Transformers

This paper introduces ClusTEX, a novel single-step graph transformer that utilizes a geometry-aware positional encoding scheme to outperform standard algorithms and multi-step deep learning approaches in reconstructing overlapping electromagnetic showers, thereby improving energy resolution, reducing splitting rates, and maintaining robustness against detector failures in both toy and ECAL-inspired calorimeter simulations.

The Gist

Imagine you are standing in a massive, dark warehouse filled with thousands of tiny, glowing lightbulbs. These lightbulbs are arranged in a giant grid. Suddenly, a beam of light (a particle) shoots into the warehouse and hits the floor. When it hits, it doesn't just make one bulb glow; it creates a cascade, a "shower" of light that ripples out, lighting up a cluster of bulbs around the impact point.

In the world of particle physics (like at the Large Hadron Collider), scientists need to figure out exactly where the beam hit and how much energy it had, just by looking at which lightbulbs are glowing and how bright they are.

The Problem: The "Crowded Room" Mess

Usually, this is easy if only one beam comes in. But in modern experiments, things get chaotic:

1. The Pileup: Imagine thousands of beams hitting the warehouse at the exact same time. The light showers from different beams overlap, creating a giant, confusing mess of glowing bulbs.
2. The Twins: Sometimes, a single particle splits into two (like a neutral pion decaying into two photons). These two new particles hit the floor so close together that their light showers merge into one big blob.
3. Broken Bulbs: Some lightbulbs in the warehouse are dead or broken, leaving dark spots in the middle of a glowing cluster.

The old way of solving this (called PFClustering) is like a very strict, rule-based security guard. The guard looks for the brightest bulb, claims it as the "center," and then grabs all the neighbors within a certain distance.

• The Flaw: If two beams overlap, the guard gets confused. He might think one big blob is two separate things (splitting the truth) or miss one entirely. If the bulbs are broken, he can't see the pattern and gives up.

The New Solution: The "Smart Detective" (Transformers)

The authors of this paper built a new kind of "Smart Detective" using Deep Learning (specifically, a type of AI called Transformers). Instead of following rigid rules, this AI learns to look at the whole picture and understand the relationships between the lights.

They built two versions of this detective:

1. The Two-Step Detective (The "Scout and Analyst" Team)

This team works in two phases:

• The Scout (SeedFinder): First, a fast AI scans the warehouse and points out the most likely spots where a beam hit. It filters out the noise.
• The Analyst (PoEN): Then, a second AI looks at the top candidates.
  • Old Analyst: Used to just look at how far apart the candidates were. If two were close, it assumed they were related. This sometimes caused it to hallucinate fake particles.
  • New Analyst (Attention-Based): This one uses Attention. Imagine the analyst has a superpower: it can "focus" on specific bulbs and ask, "Does this specific light pattern belong with that one?" It learns to ignore lights that are close together but don't make sense as a pair. This stops it from splitting one particle into two fake ones.

2. The One-Step Detective (ClusTEX)

This is the "Super Detective." It doesn't need a scout. It looks at the entire messy scene at once and solves the puzzle in a single glance.

• The Secret Sauce (Positional Encoding): This is the most creative part. The AI needs to know two things:

  1. Local Position: "Where am I relative to the center of this specific cluster?" (Like saying, "I am the 3rd bulb to the right of the brightest one.")
  2. Global Position: "Where am I in the entire warehouse?" (Like saying, "I am in the North-East corner of the building.")

  The AI combines these two. It knows that a light pattern in the North-East corner might behave differently than the same pattern in the South-West corner due to the building's shape or broken wires. This helps it handle broken bulbs and weird angles perfectly.

Why This Matters (The Results)

The authors tested their new detectives in two scenarios:

1. The "Toy" Warehouse: A simple, perfect grid.
2. The "Real" Warehouse: A complex grid with tilted angles, broken bulbs, and overlapping beams (mimicking the real CMS detector at CERN).

The Results:

• Better at Twins: When two particles hit close together (like the twins in a π⁰ decay), the old guard failed. The new AI could still separate them and measure them accurately.
• No More Splitting: The old AI often thought one particle was two. The new AI rarely makes this mistake.
• Broken Bulb Resilience: If 1% of the lightbulbs were dead, the old guard's measurements went haywire. The new AI (ClusTEX) looked at the neighbors and the global map to "guess" the missing energy, keeping its measurements stable.

The Big Picture

Think of this paper as upgrading from a rulebook to a brain.

• Old Way: "If a light is bright, grab it and its neighbors." (Fails in crowds).
• New Way: "I see a pattern here. I know how light behaves in this corner of the room. Even if some bulbs are broken, I can reconstruct the story of what happened."

This new method is a huge step forward for future particle physics experiments (like the High-Luminosity LHC), where the "warehouse" will be even more crowded and chaotic than ever before. It ensures that even in the most chaotic collisions, scientists can still find the hidden signals of new physics.

Technical Summary

1. Problem Statement

In modern hadron collider experiments (specifically the High-Luminosity LHC era), reconstructing photons and electrons is increasingly difficult due to:

• High Pileup: A large number of simultaneous collisions leads to high detector occupancy.
• Overlapping Showers: Boosted topologies (e.g., $\pi^0 \to \gamma\gamma$) and converted photons create overlapping electromagnetic showers that standard algorithms struggle to separate.
• Detector Non-uniformities: Geometric effects (e.g., crystal tilts, $(\eta, \phi)$ dependencies) and localized channel failures (dead crystals) introduce ambiguities in seed assignment and energy reconstruction.
• Limitations of Current Methods: The standard CMS algorithm, PFClustering, relies on local maxima and fixed thresholds. It suffers from:
  • High "splitting rates" (reconstructing one photon as two clusters).
  • Poor energy resolution in dense environments.
  • Inefficiency in separating boosted di-photon decays.
  • Inability to easily adapt to complex detector geometries without ad-hoc multi-pass inference.

2. Methodology

The authors propose deep learning-based approaches that reconstruct particle energy and impact position directly from calorimeter readout, moving beyond simple clustering to joint inference. The study compares two main strategies:

A. Two-Step Strategy (Updated DeepCluster)

This approach separates candidate selection from kinematic reconstruction:

1. SeedFinder: A Convolutional Neural Network (CNN) identifies candidate seed windows (7×7 crystals) likely to contain the true photon impact point.
2. Regression Network (PoEN): A Graph Neural Network (GNN) processes the selected candidates jointly to predict energy and position. Two variants are tested:
   • DW-PoEN: Uses Distance-Weighted Message Passing (fixed weights based on Euclidean distance).
   • GAT-PoEN: Uses Graph Attention Networks (GAT). This replaces fixed distance weights with learnable attention mechanisms, allowing the model to adaptively weigh information exchange between candidates based on their content rather than just proximity. This reduces spurious "splitting" of single photons.

B. Single-Step Strategy: ClusTEX

To overcome the limitations of two-step pipelines (e.g., fixed candidate counts, loss of context), the authors introduce ClusTEX, a single-step Graph Transformer.

• Architecture: It forms a local graph around a seed candidate and performs candidate selection and kinematic reconstruction in a single inference stage.
• Novel Positional Encoding: A key innovation is a dual encoding scheme:
  • Local Position: Encoded via concatenation relative to the anchor seed within the graph window. This captures the relative structure of overlapping showers.
  • Global Position: Encoded via summation relative to the detector center (absolute $\eta, \phi$). This allows the model to learn detector-specific non-uniformities and geometry-dependent effects.
• Graph Formation: Candidates are dynamically grouped into local graphs (approx. 13×13 effective window) based on energy thresholds, handling variable occupancy without fixed window constraints.

3. Key Contributions

1. Attention Mechanism for Clustering: Demonstrated that replacing distance-driven message passing with attention mechanisms significantly reduces "splitting" (false double-reconstruction) and improves energy resolution for overlapping showers, eliminating the need for ad-hoc multi-pass inference.
2. ClusTEX Architecture: Introduced a unified single-step transformer that outperforms two-step methods by jointly learning selection and reconstruction, offering superior robustness in complex topologies.
3. Hybrid Positional Encoding: Developed a novel encoding strategy that separates local graph topology from global detector geometry, enabling the model to handle $(\eta, \phi)$ dependencies and localized detector failures effectively.
4. Robustness to Failures: Showed that ClusTEX can partially compensate for non-responsive channels (dead crystals) by leveraging contextual information from neighboring active crystals and global position awareness.

4. Results

The models were trained and evaluated on Geant4 simulations of two setups: a simplified "toy" calorimeter and a realistic CMS ECAL-inspired topology (including crystal tilts, material upstream, and $(\eta, \phi)$ dependencies).

• Toy Calorimeter (Overlapping Showers):

  • Splitting: The attention-based GAT-PoEN reduced the splitting rate to 0.05% (vs. 0.56% for GNN and standard for PFClustering).
  • Resolution: Both ML models significantly outperformed PFClustering in energy resolution ($\sigma_E$) for 2-photon samples.
  • Boosted $\pi^0$: In boosted $\pi^0 \to \gamma\gamma$ events ($p_T > 30$ GeV), PFClustering became inefficient, while the attention-based model retained high reconstruction capability and mass resolution.

• ECAL-Inspired Topology (Realistic Conditions):

  • Performance: ClusTEX achieved the best overall performance, yielding improved energy resolution and the lowest splitting rate (0.06% for 2-photon samples) compared to two-step approaches and PFClustering.
  • Background Rejection: ClusTEX showed superior background rejection compared to two-step methods.
  • Robustness: Under simulated conditions with 1% disabled crystals and dead readout towers, ClusTEX maintained stable energy reconstruction with reduced bias, whereas PFClustering and two-step models showed significant distortions.

5. Significance

This work represents a significant step forward in calorimeter reconstruction for high-luminosity environments:

• Scalability: The single-step ClusTEX approach offers a scalable foundation for future high-granularity detectors (like the CMS HGCAL) where traditional clustering algorithms struggle with complexity.
• Physics Impact: By improving the separation of overlapping showers, the method enhances the reconstruction of boosted objects (e.g., $\pi^0$, $\eta$, heavy flavor decays), which is critical for precision Standard Model measurements and searches for new physics.
• Detector Resilience: The ability to recover energy in the presence of localized detector failures suggests these models can maintain performance even as detectors age or suffer hardware issues, reducing the need for complex calibration corrections.
• Foundation for Future Work: The proposed methodology serves as a "foundation" for downstream tasks like superclustering and particle flow, potentially unifying the reconstruction chain from raw hits to final physics objects.