⚛️ high-energy experiments

Optimised Graph Convolution for Calorimetry Event Classification

This paper presents an optimized graph convolution framework, featuring efficient graph construction and specialized layers, to address particle identification and energy regression challenges in high-granularity calorimeters within resource-constrained environments.

Original authors: Matthieu Melennec, Shamik Ghosh, Frédéric Magniette

Published 2026-07-27

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Matthieu Melennec, Shamik Ghosh, Frédéric Magniette

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to understand a massive, chaotic city by looking at it from a helicopter. If you try to look at every single brick, every car, and every person all at once, your brain would explode. This is the problem facing modern particle physics. Scientists are building giant, super-detailed detectors to catch tiny, invisible particles zipping through space. These detectors are like cities made of millions of tiny sensors, each one shouting out data when a particle bumps into it. The challenge is that there is so much data, and the patterns are so messy, that old-school computer programs get lost in the noise. They are like trying to find a specific friend in a crowded stadium by looking at a list of names rather than seeing where people are standing. To solve this, scientists are turning to a type of artificial intelligence called Graph Neural Networks. Think of these not as a rigid grid (like a spreadsheet), but as a dynamic map of connections. Instead of forcing data into neat rows and columns, these networks treat the data like a social network, where every sensor is a person, and the connections between them are friendships. By studying how these "friends" talk to each other, the computer can figure out who the mysterious visitor is and how much energy they brought to the party.

The paper you are about to read, titled "Optimised Graph Convolution for Calorimetry Event Classification," is a recipe for making this social-network-style AI faster and smarter, specifically for a high-tech detector that looks a bit like a giant, layered honeycomb. The authors, working at a French research lab, wanted to teach a computer to look at the "shower" of energy left behind by a particle and say, "Ah, that was an electron!" or "That was a pion!" and also guess exactly how much energy it had.

The main trick they used is a clever shortcut for building these social networks. Usually, to build a graph, a computer has to check every single sensor against every other sensor to see who is close enough to be a friend. This is like asking every person in a stadium to check their phone against every other person's phone to find their neighbors; it takes forever and uses up all the battery. The authors invented a new method called "PT-KNN" (Proximity Table k-Nearest Neighbors). Imagine that before the party even starts, the organizers have already written down a pre-sorted list of who is standing next to whom for every single seat in the stadium. When the computer needs to build a graph, it doesn't have to do the hard math of measuring distances; it just looks up the pre-made list. This makes the process much faster and uses less memory, which is crucial if you want to run this on a small, power-hungry chip inside the detector itself.

Once the graph is built, the computer uses "message passing." This is like a game of telephone where sensors pass notes to their neighbors. A sensor says, "Hey, I felt a big hit!" and its neighbors pass that message along, updating their own understanding of the situation. The authors also introduced a new way to "pool" or shrink the graph, which they call "Treclus." Imagine the graph is a detailed map of a city. As the computer gets smarter, it doesn't need to see every single street; it can zoom out and see neighborhoods. Treclus groups nearby sensors together into clusters, simplifying the map without losing the big picture of the particle's path. Finally, they have a special "readout" step that flattens this complex 3D map into a simple list of numbers that a standard computer brain (a Multi-Layer Perceptron) can easily read to make a final decision.

The results of their simulations are quite promising. They tested their system on a simulated dataset of four types of particles: electrons, photons, muons, and pions, with energies ranging from 10 to 100 GeV. The system was very good at telling them apart. For example, it could distinguish between an electron and a pion with high accuracy, only getting confused when a particle did something weird, like a pion that started acting like an electron too early. When it came to guessing the energy, the system was also impressive. For electrons and photons, the error in their energy guess was about 20%, which matches what scientists expect from the laws of physics. This suggests that their method of using pre-computed lists and smart clustering works just as well as the slower, more traditional methods, but with a fraction of the computing power.

In short, this paper shows that by being smart about how we organize the data—using pre-made lists instead of doing math on the fly, and grouping sensors into neighborhoods—we can teach computers to understand the chaotic world of particle collisions much more efficiently. It's a step toward making the next generation of particle detectors not just smarter, but also faster and more energy-efficient, ready to handle the massive amounts of data coming from future experiments.

Technical Summary: Optimised Graph Convolution for Calorimetry Event Classification

Problem Statement
The increasing luminosity of the High-Luminosity Large Hadron Collider (HL-LHC) and the enhanced granularity of detectors like the CMS HGCAL upgrade have led to a massive increase in data volume and channel count. While this improves the precision of observing complex phenomena and rare events, it renders standard reconstruction and triggering techniques ineffective. Specifically, energy threshold-based triggers struggle with high pile-up complexity, and standard deep learning methods (e.g., convolutional neural networks) fail to generalize well to the peculiar, non-grid topologies of particle detectors. There is a critical need for algorithms that can efficiently process high-granularity calorimeter data while respecting the geometric and physical symmetries of the detector.

Methodology
The authors propose a Graph Convolutional Network (GCN) framework designed for particle identification (PID) and energy regression in high-granularity calorimeters. The pipeline consists of four main stages:

Optimised Graph Construction (PT-KNN):
- Events are represented as graphs where nodes correspond to hit sensor cells and edges represent geometric proximity.
- Instead of computing Euclidean distances dynamically for every event (which is computationally expensive), the authors introduce Proximity Tables (PTs). These tables pre-compute the order of neighboring sensors for every detector cell based on a user-defined metric (Euclidean distance in this study).
- The PT-KNN algorithm utilizes these tables to construct graphs by selecting the $k$ nearest neighbors from the pre-sorted lists. This approach significantly reduces time complexity compared to standard KNN, achieving an average complexity of $O(N \langle c_{last} \rangle) \sim (\log N)^2$ in practice, and linear asymptotic complexity $O(N)$ under specific conditions. This is crucial for resource-constrained environments like hardware triggers.
Message Passing Convolution:
- The framework employs message passing where node features (energy) and edge features (Euclidean distance) are updated.
- A message function (single neuron layer with Leaky-ReLU) processes concatenated node and edge features.
- Aggregation strategies differ by task: Max pooling is used for particle identification to highlight protruding features, while Mean pooling is used for energy regression to smooth features and consider the graph globally. Self-loops are added to preserve node information during updates.
Geometric Pooling (Treclus):
- To coarsen the graph while preserving global structure, the authors introduce Treclus, a threshold-based clustering algorithm.
- Treclus creates a matching of nodes linked by edges shorter than a threshold $\epsilon$ . To maintain linear complexity and avoid the quadratic problem of finding a maximum matching, the algorithm is applied iteratively (less than 10 calls suffice).
- Clusters are pooled via feature-wise operations (max for PID, sum for energy regression to conserve node count information). Cluster positions are randomly selected from constituent nodes to maintain embedding within the known detector geometry.
Graph Readout and MLP:
- To feed the variable-sized graph output into a Multi-Layer Perceptron (MLP), a Readout Pooling layer flattens the graph into a fixed-size tensor.
- Crucially, this step respects the detector's rotational symmetry by dividing the detector into specific readout regions (concentric rings and longitudinal slices) based on the physics of the interaction (e.g., distinguishing electromagnetic vs. hadronic showers). This prevents the model from learning redundant representations for rotated identical showers.
- The final MLP is funnel-shaped with 6 layers and Leaky-ReLU activations.

Key Contributions

PT-KNN Algorithm: An optimized graph construction method using pre-computed proximity tables that drastically reduces the computational cost of graph generation, making it viable for resource-constrained environments (e.g., FPGAs).
Treclus Pooling: A novel clustering algorithm designed to coarsen graphs based on geometric locality while preserving the overall graph structure and maintaining linear complexity.
Symmetry-Aware Readout: A readout strategy that explicitly incorporates detector geometry and physical symmetries (rotational and longitudinal) to flatten graph data, ensuring the model does not need to learn redundant symmetries.
Integrated Pipeline: A complete framework demonstrating both particle classification and energy regression for electrons, photons, muons, and pions in a high-granularity calorimeter simulation.

Results

Classification: The model achieves high accuracy in distinguishing between muons, pions, electrons, and photons. Misclassifications are primarily attributed to outlier events, such as electromagnetic showers leaking into the hadronic calorimeter or early showering pions.
Energy Regression: The energy response ( $E_{pred}/E_{true}$ ) follows a Gaussian-like distribution centered around 1.
Resolution: The energy resolution for electromagnetic showers follows the expected stochastic behavior $\sigma/\mu \propto 1/\sqrt{E}$ . The stochastic term for the electromagnetic calorimeter was measured at $S_{GNN} = 20.15\%$ , closely matching the theoretical value of 20%.
Efficiency: The use of proximity tables and graph truncation (exploiting only ~5% of the proximity table rows in practice) significantly reduces memory footprint and computational load without sacrificing graph quality.

Significance and Claims
The paper claims that this framework provides a state-of-the-art solution for high-granularity calorimeter data processing. By leveraging the specific geometry of the detector through Proximity Tables and symmetry-aware readout, the approach overcomes the limitations of standard CNNs and traditional reconstruction techniques. The authors assert that their method is detector-agnostic and can be generalized to other non-regular high-granularity detector challenges, such as diffuse supernova neutrino background searches in Super/Hyper-Kamiokande or overlapping object segmentation. The successful recovery of expected energy resolution values validates the physical correctness of the approach. Furthermore, the optimization of graph construction and the potential for truncating proximity tables suggest the method is suitable for deployment in real-time triggering systems with strict resource constraints.

Technical Summary: Optimised Graph Convolution for Calorimetry Event Classification

More like this