⚛️ high-energy experiments

Comprehensive reconstruction of collider events with hypergraph representation learning and graph-conditioned diffusion

This paper introduces VyPER, a novel geometric learning framework that combines hypergraph representation learning with graph-conditioned diffusion models to simultaneously optimize particle assignment and neutrino kinematics prediction, demonstrating superior event reconstruction performance across diverse Standard Model processes compared to existing techniques.

Original authors: Lining Mao, Yvonne Peters, Ethan Simpson, Zihan Zhang

Published 2026-09-17

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Lining Mao, Yvonne Peters, Ethan Simpson, Zihan Zhang

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

In the heart of Europe, inside a massive ring of magnets buried beneath the border of France and Switzerland, protons are smashed together at nearly the speed of light. These collisions are the most energetic events humans can create, momentarily recreating conditions similar to those just after the Big Bang. When these protons collide, they shatter into a shower of new, short-lived particles that decay almost instantly into stable debris. Detectors surrounding the collision point act like giant, ultra-fast cameras, recording the paths and energies of these stable fragments, such as jets of particles and charged electrons or muons. However, the most interesting players in these collisions—heavy particles like the top quark or the Higgs boson—disappear before they can be seen directly. Scientists must act as cosmic detectives, trying to reconstruct the invisible parents from the visible children they left behind. The challenge is immense because some of these decay products are neutrinos, ghostly particles that pass through the detectors without leaving a trace, carrying away energy and momentum that must be inferred rather than measured.

For decades, physicists have relied on complex mathematical formulas and rigid rules to piece these events back together. These traditional methods work well when the physics is simple and the equations can be solved exactly. But when the collision produces multiple invisible particles or when the data is messy, these formulas often break down, leaving scientists with impossible equations or multiple conflicting answers. A new study introduces a different approach, one that treats the entire collision event not as a set of equations to be solved, but as a connected network of relationships to be learned. The researchers developed a system called VyPER, which uses a type of artificial intelligence designed to understand complex structures. Instead of forcing the data into a pre-defined box, VyPER looks at the collision as a web of connections, learning how the visible particles relate to each other and how they might have originated from invisible parents.

The core of this new method involves a clever way of representing the data. Imagine the particles in a collision as dots on a map, connected by lines that show how they are related. In this system, the invisible neutrinos are included in the map even though they were never seen, linked only to the particles they are known to share a parent with. The system then uses a process similar to how a group of people might share information to solve a puzzle. Each particle passes details about its energy and direction to its neighbors, and through many rounds of this exchange, the system builds a deep understanding of the entire event. This allows the computer to figure out which visible particles came from which parent, a task known as assignment, and to predict the missing momentum of the invisible neutrinos.

To test if this approach works, the researchers simulated four different types of proton collisions that occur at the Large Hadron Collider. These included the production of top quark pairs, Higgs bosons decaying into W bosons, and rare, complex events involving both top quarks and W bosons. In every case, they compared the new system against the best existing methods, including traditional mathematical solvers and other machine learning models. The results showed that the new system could correctly identify which particles belonged together more often than the old methods. More importantly, when it came to guessing the path and energy of the invisible neutrinos, the new system provided answers that were closer to the true values than the traditional techniques.

The study found that while the traditional mathematical methods are powerful, they struggle when the data is imperfect or when there are too many unknowns. These older methods often fail to find a solution for a significant number of events, or they produce answers that are mathematically possible but physically unlikely. The new system, by contrast, learned the patterns of the collisions from the simulation data itself. It did not rely on rigid rules but instead learned the probability of different outcomes. In the tests, it successfully reconstructed the invisible parts of the collision for nearly every event, whereas the traditional methods left many events unsolved. When the researchers looked at the final reconstructed properties of the parent particles, such as their mass and speed, the new system's results matched the true values more closely than the alternatives.

This success is particularly notable in the most difficult scenarios, where the collision produces multiple invisible particles and the math becomes too tangled for standard formulas to untangle. In these cases, the new system demonstrated an ability to learn the hidden constraints of physics that are not explicitly written in the equations. It effectively learned to navigate the "under-determined" nature of the problem, where there are more unknowns than equations, by using the relationships between all the visible particles to guide its guesses. The researchers showed that this approach works across a wide variety of collision types, suggesting it could become a standard tool for analyzing future data from particle colliders.

The implications of this work extend beyond just getting better numbers. By accurately reconstructing the full picture of a collision, including the invisible parts, physicists can measure the properties of fundamental particles with greater precision. This could help them spot tiny deviations from the known laws of physics, which might be the first signs of new, undiscovered particles or forces. The study confirms that machine learning, when designed with a deep understanding of how particles interact, can solve problems that have long been considered too difficult for computers. It does not replace the laws of physics but rather provides a more flexible and powerful way to apply them to the messy reality of experimental data.

In the end, the researchers have demonstrated that a new way of thinking about particle collisions—one that treats them as interconnected networks rather than isolated equations—can unlock a clearer view of the subatomic world. The system they built, VyPER, successfully combined the task of sorting particles with the task of predicting the invisible, doing both at once with a level of accuracy that surpassed current standards. While the results come from computer simulations, they offer a strong proof that this method is ready to be tested on real data from the Large Hadron Collider. If it performs as well on real collisions as it did in the simulation, it could open new avenues for precision measurements in the study of the Higgs boson, the top quark, and the electroweak force, helping scientists to peer deeper into the fundamental structure of the universe.

Technical Summary: Comprehensive Reconstruction of Collider Events with Hypergraph Representation Learning and Graph-Conditioned Diffusion

Problem Statement
In particle collider experiments, such as those at the Large Hadron Collider (LHC), event reconstruction involves inferring the kinematics of short-lived parent particles (e.g., top quarks, W/Z bosons, Higgs bosons) from the stable final states recorded by detectors. This task is decomposed into two primary challenges:

Final-State Assignment: Correctly associating measured jets and charged leptons to their respective parent particles. This is a combinatorial problem where incorrect pairings lead to erroneous kinematic reconstructions.
Neutrino Prediction: Inferring the kinematics of unmeasured neutrinos. This is an ill-posed inverse problem because detectors cannot directly observe neutrinos; only missing transverse momentum ( $E_T^{miss}$ ) is measured. The underlying distribution of neutrino kinematics is often non-Gaussian and multi-modal, and traditional analytical methods (e.g., $\chi^2$ minimization or kinematic ellipse methods) frequently fail to produce physical solutions due to detector smearing or insufficient kinematic constraints.

Existing machine learning (ML) approaches have typically addressed these tasks in isolation or for specific physics processes (e.g., only $t\bar{t}$ or only $H \to WW$ ), limiting their generalizability across the Standard Model (SM).

Methodology: The VyPER Framework
The authors present VyPER, a unified geometric learning framework designed to perform comprehensive event reconstruction across diverse proton-proton collision processes. VyPER integrates supervised classification for particle assignment with a generative diffusion model for neutrino prediction within a single architecture.

Hypergraph Representation:
Collider events are modeled as directed heterogeneous graphs $G=(V, E)$ .
- Nodes ( $V$ ): Represent experimentally detectable final states ( $V_{exp}$ , e.g., jets, leptons) and invisible final states ( $V_{inv}$ , i.e., neutrinos).
- Edges ( $E$ ): Represent connections between nodes. Crucially, neutrinos are connected only to their "companion" visible particles (e.g., a lepton from the same $W$ decay) via "conditional edges."
- Hyperedges ( $X$ ): To capture higher-order correlations (e.g., a top quark decaying into three partons), the graph is extended to a hypergraph where hyperedges connect three or more nodes.
Message-Passing and Latent Construction:
The backbone utilizes a Message-Passing Neural Network (MPNN). A novel "conditional message-passing" mechanism is introduced:
1. The network aggregates information from visible nodes to construct a latent representation ( $\tilde{x}$ ) for each neutrino.
2. These neutrino latents are propagated back to the graph to modulate the features of their companion nodes (using Adaptive Layer Normalization, AdaLN).
3. This process iterates $S$ times, updating edge, node, and global features to capture kinematic relations.
Task Heads:
1. Assignment (Classification): The network performs edge classification (for two-body decays) and hyperedge classification (for three-body decays) to identify the correct grouping of final states belonging to a parent particle.
2. Neutrino Prediction (Diffusion): A diffusion head predicts neutrino kinematics. It employs a Denoising Diffusion Implicit Model (DDIM) conditioned on the learned neutrino latent representations. Instead of regressing a single value, the model learns the conditional probability density of neutrino momenta, allowing it to sample multiple plausible solutions.
Unified Loss Function:
The model is trained end-to-end using a joint loss function combining the diffusion loss (MSE between predicted and true noise) and cross-entropy losses for edge and hyperedge classification, weighted by hyperparameters $\eta$ and $\alpha$ .

Experimental Setup
The framework was evaluated on four distinct Standard Model processes using simulated data at $\sqrt{s} = 13$ TeV:

$t\bar{t} \to 2L$ : Dileptonic top-antitop production (benchmark against analytical methods).
$H \to WW^*$ : Higgs boson production via vector-boson fusion decaying to $WW$ .
VBS $WW$ : Vector-boson scattering producing same-sign $WW$ .
$t\bar{t}W$ : Associated production of top-antitop and a $W$ boson (highly complex, multi-neutrino topology).

Performance was compared against state-of-the-art ML models (SPANet, $\nu^2$ -flows) and traditional analytical solvers (Ellipse Method). Metrics included assignment efficiency, neutrino momentum residuals (RMSE, $\Delta R$ ), and the fidelity of reconstructed high-level observables (invariant masses, angular distributions) against truth-level distributions.

Key Results

Assignment Performance: VyPER achieved assignment efficiencies comparable to or slightly better than SPANet across all processes (e.g., ~85% for $t\bar{t}$ leptonic top assignment).
Neutrino Reconstruction:
- In the $t\bar{t}$ channel, VyPER and $\nu^2$ -flows outperformed the analytical Ellipse Method (EM) in terms of resolution and the fraction of events yielding physical solutions.
- In under-constrained electroweak channels ( $H \to WW^*$ , VBS $WW$ ), VyPER demonstrated superior accuracy in neutrino kinematics compared to $\nu^2$ -flows, particularly in the VBS $WW$ process where it reduced the $\Delta R$ separation and momentum residuals significantly.
High-Level Observables: VyPER successfully recovered the shapes of complex distributions (e.g., the double-peak structure of $W$ boson masses in $H \to WW^*$ ) and showed higher diagonal purity in migration matrices for observables like $m_{t\bar{t}}$ and $m_{WW}$ compared to competing methods.
Complex Topologies: In the $t\bar{t}W$ process, VyPER successfully handled the simultaneous assignment of jets to hadronic tops, leptons to leptonic tops, and the prompt $W$ , outperforming SPANet in reconstruction efficiency for all system components.

Significance and Claims
The paper claims that VyPER demonstrates that accurate, comprehensive event reconstruction is achievable across a diverse range of SM physics processes using a unified ML framework. Key contributions include:

Generalizability: Unlike previous models restricted to specific topologies, VyPER is designed to be applicable to arbitrary physics processes by leveraging hypergraph representations.
Integration of Tasks: It successfully unifies the combinatorial assignment problem and the generative neutrino prediction problem, showing that learning these tasks jointly improves overall performance.
Handling Under-Constrained Systems: The framework proves capable of resolving kinematically under-constrained systems (like VBS $WW$ ) where traditional analytical methods fail or require strong physical assumptions.
Precision Physics: The authors suggest that these techniques open new avenues for precision measurements in the Higgs, electroweak, and top-quark sectors, potentially enhancing the separation of signal from background and improving the accuracy of unfolded differential cross-sections.

The paper concludes that while analytical methods remain powerful when combinatorial ambiguities are resolved, generative ML approaches like VyPER offer a robust alternative for complex, high-multiplicity final states where analytical solutions are intractable.

More like this