🔬 physics

Denoising Graph Super-Resolution towards Improved Collider Event Reconstruction

This study demonstrates that integrating a novel transformer-based super-resolution model into an LHC-like reconstruction pipeline effectively enhances calorimeter granularity and suppresses noise, significantly improving particle reconstruction quality as a cost-effective complement to high-granularity hardware detectors.

Original authors: Nilotpal Kakati, Etienne Dreyer, Eilam Gross

Published 2026-10-30

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Nilotpal Kakati, Etienne Dreyer, Eilam Gross

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

High-energy physics is the science of smashing particles together at incredible speeds to see what they are made of and how they behave. To do this, scientists build massive machines called colliders, where beams of protons or electrons crash into one another. When these particles collide, they shatter into a spray of new, often short-lived particles. To understand what happened, detectors surrounding the collision point act like giant, ultra-sensitive cameras, capturing the energy and path of every fragment. One of the most important tools in these detectors is the calorimeter, a thick block of dense material designed to stop particles and measure their energy. As a particle crashes into the calorimeter, it creates a shower of secondary particles, spreading its energy out across many small sections, or cells, of the detector. The more finely these cells are cut, the clearer the picture scientists can build of the original collision. However, building a detector with extremely tiny cells is incredibly expensive and mechanically difficult. This creates a constant tension: scientists want the sharpest possible images of the subatomic world, but they are often limited by the cost and complexity of the hardware.

In a new study, researchers at the Weizmann Institute of Science in Israel explored a clever workaround to this problem. Instead of building a more expensive detector, they asked if they could use advanced computer software to make a less detailed detector look like a more detailed one. This technique, known as superresolution, is already used in fields like photography to turn blurry, low-quality images into sharp, high-definition ones. The team applied this idea to the chaotic, noisy data coming from a particle detector. They developed a software system that takes the coarse, low-resolution energy readings from a simulated detector and uses a neural network to predict what the data would have looked like if the detector had been built with much finer cells. The goal was not just to make the image prettier, but to see if this "software upgrade" could actually help scientists reconstruct the original particles more accurately, especially in crowded environments where many particles arrive at once.

The researchers tested their idea using a simulated version of a detector similar to those used at the Large Hadron Collider. They created two versions of the same data: a high-resolution version that represented the ideal, fine-grained detector, and a low-resolution version that mimicked a cheaper, coarser detector with added electronic noise. They then fed the noisy, low-resolution data into their new software model. The model, which is built on a mathematical framework called a continuous normalizing flow, learned to reverse the blurring and noise. It took the coarse data and generated a new, high-resolution map of energy deposits. Crucially, the model was trained to suppress the random electronic noise that often distorts measurements, while simultaneously filling in the missing details between the coarse cells. The result was a clean, detailed image of the particle shower that looked very similar to the ideal high-resolution truth, even though it started from a much blurrier input.

To see if this software enhancement actually helped with the real work of physics, the team plugged the improved data into a particle-flow algorithm. This is a standard method used to identify and measure the properties of the particles created in a collision, such as their energy, direction, and type. They compared how well this algorithm performed when fed the original low-resolution data versus the new, super-resolved data. The results were clear: the software-enhanced data led to significantly better reconstructions. The algorithm could distinguish between particles that were very close together much more effectively, and it measured their energies with greater precision. In tests involving single electrons, the software reduced the error in energy measurement and successfully removed the noise that usually skews the results. When they moved to more complex scenarios with multiple particles, the improvement was even more striking. The model trained on the super-resolved data was much better at counting how many particles were present and determining their exact paths, narrowing the gap between what a cheap detector could achieve and what a perfect, expensive one could do.

The researchers were careful to ensure that the software was not just inventing details or "hallucinating" features that weren't there. They found that while the model could not recover the exact, specific truth of a single event—which is impossible because information is lost when the detector is coarse—it could generate a representation that was physically consistent with the laws of physics. The software learned the statistical patterns of how particles shower through a detector and used that knowledge to fill in the gaps. This means the output is not a random guess, but a highly educated reconstruction that respects the underlying physics. The team also tested whether the improvement came simply because the software model was larger or more complex, or if it truly came from the super-resolution step. They found that even when they made the particle-flow algorithm itself much larger, it still could not match the performance of the smaller algorithm fed with the super-resolved data. This proved that the gain came from the quality of the input data provided by the super-resolution step, not just from having a bigger computer program.

One of the most important findings was that the software did not just improve the total energy measurement, which is a simple sum of all the parts. It also dramatically improved the ability to pinpoint the direction of the particles. In a low-resolution detector, a particle's path often gets pinned to the center of a single large cell, creating a "staircase" effect where the direction jumps in big steps. The super-resolution model smoothed this out, allowing the reconstruction to follow the true path of the particle with much finer precision. This is critical for modern physics, where scientists need to distinguish between particles that are flying in nearly the same direction. The study showed that this software approach could effectively narrow the performance gap between a coarse detector and a fine one, offering a powerful tool for future experiments. While the researchers noted that this method relies on simulations and would need rigorous testing with real data to handle the complexities of actual collisions, the results serve as a strong proof of concept. They demonstrate that with the right software, we might be able to get the clarity of a million-dollar detector from a much more affordable one, opening new doors for exploring the fundamental nature of the universe without the need for impossible engineering feats.

Technical Summary: Denoising Graph Superresolution toward Improved Collider Event Reconstruction

Problem Statement
Future high-energy physics facilities, such as Higgs factories and energy-frontier colliders, require high-granularity calorimeters to accurately reconstruct particle jets and distinguish nearby particles in dense environments (e.g., high-luminosity LHC pileup). However, the cost and engineering complexity of building detectors with extremely fine segmentation are substantial. Low-granularity calorimeters suffer from limited imaging capabilities, constraining the performance of Particle-Flow (PFlow) algorithms in tasks such as assigning energy deposits to tracks, identifying neutral hadrons, and separating nearby particles. While software-based approaches like superresolution (SR) have been explored in image processing, their integration into a full collider reconstruction pipeline to enhance calorimeter granularity and suppress noise remains an open challenge.

Methodology
The authors propose a framework that integrates superresolution techniques directly into the core pipeline of particle reconstruction. The methodology consists of two primary components:

Graph-Based Superresolution Model:
- Data Representation: Calorimeter data is represented as graphs to handle irregular geometry, nontrivial spatial correlations, and signal sparsity. Low-resolution (LR) cells are expanded into high-resolution (HR) graphs where each LR cell is subdivided into $f_\eta \times f_\phi$ HR cells.
- Generative Framework: The model employs Continuous Normalizing Flows (CNF) trained via Flow Matching. This generative approach transforms a simple base distribution into a complex target distribution (the truth HR energy deposits) through a continuous sequence of invertible transformations governed by a neural ordinary differential equation.
- Denoising and Conditioning: The model is conditioned on the LR energy deposits (upsampled by repetition) to predict the HR energy values. It simultaneously learns to suppress electronic noise by training on simulated data where noise is applied to the LR configuration but absent in the HR truth.
- Ensemble Sampling: To mitigate the stochastic nature of the generative process and approximate the deterministic target, the authors employ an ensemble approach, averaging the outputs of multiple forward passes through the same model.
Novel Transformer-Based Particle-Flow (PFlow–) Model:
- To evaluate the impact of SR on downstream reconstruction, the authors introduce a new PFlow architecture (PFlow–) that replaces iterative learning with a supervised attention mechanism.
- Architecture: The model uses a transformer network to generate neighborhood-aware representations of detector cells. It branches into two paths: one predicting particle cardinality (via classification) and another predicting kinematics.
- Incidence Matrix: Instead of treating reconstruction as a black-box regression, the model explicitly learns the fractional energy distribution (incidence matrix) of cells over particles using a cross-attention mechanism. This ensures interpretability, allowing direct inspection of how energy is assigned to reconstructed particles.
- Correction Network: A variant of the model includes a correction network (a residual MLP) to refine proxy kinematics, testing whether performance gains are due to the input granularity or the estimator's ability to correct for it.

Datasets
The study utilizes the COCOA package, which simulates an LHC-like generic calorimeter using Geant4. Two datasets were generated:

Single Electron: A simplified set with a single electron ( $p_T \in [50, 51]$ GeV).
Multiparticle: A complex set involving an electron and $N$ photons ( $N \in [0,3]$ ) with angular separations comparable to cell sizes.
In both cases, LR data is derived by downsampling HR truth data and adding electronic noise, while HR data serves as the ground truth.

Key Results

Energy Resolution: The SR model significantly improves energy resolution. The residual distribution of predicted HR energies is narrower and centered closer to zero compared to measured LR energies. Ensemble sampling (size 10) was found to be optimal for balancing computational cost and performance.
Substructure Recovery: The model successfully recovers substructure features obscured by coarse segmentation. Energy correlator observables ( $C_2, C_3, D_2$ ) computed on predicted HR cells closely match the truth distributions, outperforming noisy LR data.
Particle-Flow Performance:
- Cardinality: The PFlow model trained on predicted HR cells shows improved or comparable performance in identifying the number of particles, particularly as true cardinality increases.
- Angular Resolution: The model trained on HR inputs demonstrates a marked improvement in angular resolution ( $\eta$ and $\phi$ ). Without a correction network, the LR model suffers from a "staircase" effect where predictions are quantized to cell centers; the HR model avoids this.
- Decoupling Noise and Granularity: By summing predicted HR cells back to the original LR scale (retaining only the denoising benefit), the authors found that noise suppression primarily improves energy resolution, while the finer segmentation (granularity) is responsible for the gains in angular resolution.
- Robustness: Even when both models (LR and HR inputs) are equipped with a correction network to refine kinematics, the model trained on SR-predicted HR cells retains a significant performance advantage (factors of 1.5–1.9 in angular resolution and 1.6 in energy), indicating the improvement stems from the input representation rather than the estimator's form.

Significance and Claims
The paper presents a proof of concept demonstrating that superresolution is a viable software-based tool to enhance collider event reconstruction. The authors claim that:

Feasibility: SR can be effectively integrated into a particle-flow reconstruction chain, narrowing the performance gap between low-granularity detectors and finer ones, though not entirely closing it.
Interpretability: The proposed transformer-based PFlow model offers improved interpretability by explicitly modeling the energy assignment matrix, allowing for transparent analysis of how SR impacts reconstruction.
Cost-Performance Trade-off: By improving reconstruction quality at a fixed detector granularity, SR offers a pathway to optimize the performance-to-cost trade-off in future detector designs without replacing the physical measurement capabilities of a finer detector.
Limitations: The authors modestly acknowledge that SR relies on physical priors learned from simulation (Standard Model showers) and cannot recover information lost during measurement that contradicts these priors. Consequently, it is not a replacement for actual high-granularity hardware, particularly for anomalous physics scenarios (e.g., exotic particles) or in the absence of rigorous data-simulation alignment strategies.

The study concludes that while specific model architectures may evolve, the core finding—that superresolution can enhance reconstruction in collider environments—is robust and warrants further exploration in realistic, high-multiplicity collision scenarios.

Technical Summary: Denoising Graph Superresolution toward Improved Collider Event Reconstruction

More like this