⚛️ phenomenology

Down-Type Jet Identification in Fully Hadronic $t\bar t$ Events

This paper presents a novel two-stage machine-learning framework combining GNNs, Transformers, and conditional diffusion models to successfully identify down-type jets in fully hadronic $t\bar{t}$ events, thereby enabling effective spin-correlation and entanglement studies in a channel previously hindered by QCD background and reconstruction challenges.

Original authors: Jinhui Guo, Chenhao Peng

Published 2026-08-25

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Jinhui Guo, Chenhao Peng

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 subatomic world, the top quark is a giant. It is the heaviest known elementary particle, so massive that it decays into other particles before it can even form a stable structure like an atom. Because it dies so quickly, the way it spins when it is created is imprinted on the directions of the particles it leaves behind. Physicists study these directions to understand the fundamental rules of quantum mechanics, specifically how particles can be "entangled," a phenomenon where two objects remain linked in a way that defies everyday intuition. To see this entanglement clearly, scientists need to track the specific paths of the decay products with extreme precision.

For years, researchers have focused on top quarks that decay into particles including electrons or muons, which are easier to spot and measure. However, the most common way a top quark pair decays is into a shower of ordinary particles called jets, a scenario known as the "fully hadronic" channel. This path is statistically the most abundant, offering four times more events than the electron-based methods. Yet, it has been largely ignored for spin studies because it is a chaotic mess. In this channel, the top quarks produce jets that look identical to the background noise of the particle collider, and the specific particles needed to measure the spin are buried inside these jets. Without a way to sort the signal from the noise, the valuable information about the top quark's spin is lost.

A team of researchers has now developed a new method to untangle this mess, turning the most difficult channel into a viable tool for quantum study. They created a two-stage artificial intelligence system designed to reconstruct the entire event from scratch. First, the system looks at all the jets produced in a collision and sorts them into the correct groups, identifying which jets came from the heavy top quarks and which came from the lighter particles. It then performs a second, more delicate task: distinguishing between two types of particles inside those jets that look nearly identical but carry opposite information about the spin. By training this system on millions of simulated collisions, the researchers found that their new method could successfully identify the correct particle paths, recovering spin information that was previously inaccessible.

The core of the challenge lies in the nature of the decay. When a top quark decays, it produces a W boson, which then splits into two jets. One of these jets comes from a "down-type" quark, which acts as a perfect compass for the top quark's spin, while the other comes from an "up-type" quark, which provides no such directional clue. In the fully hadronic channel, these two jets are indistinguishable to standard detectors. If a physicist guesses wrong about which is which, the spin measurement becomes diluted or completely wrong. The researchers' solution was to build a neural network that doesn't just guess, but learns to recognize subtle patterns in the internal structure of the jets, such as how the energy is distributed among the tiny particles inside them.

The team tested their system using a massive dataset of simulated collisions, representing the conditions of a high-energy particle collider. They divided the work into two steps. The first step acted as a broad sorter, taking all the jets from an event and arranging them into a legal six-jet candidate that matches the expected pattern of two top quarks. This step had to filter out the overwhelming background noise of random particle collisions. The second step took the six selected jets and tried to solve the specific puzzle of the down-type versus up-type quarks. To improve this second step, they introduced a novel training technique involving a "diffusion" process, which is a method of teaching the computer to recognize patterns by gradually adding and then removing noise from data.

Surprisingly, the researchers discovered that the standard way of training this diffusion component did not work for their specific goal. When they tried to teach the system simply to remove noise, the resulting scores were no better than random guessing. However, when they changed the training goal to focus on ranking the correct answer higher than the wrong ones, the system suddenly became highly effective. This finding suggests that for this type of complex sorting problem, the goal of the training must be tailored specifically to the task of distinguishing between similar options, rather than just general pattern recognition.

The results of the simulation showed a significant improvement over previous methods. When the researchers applied their new two-stage system, they were able to identify the correct particle paths with much higher accuracy than if they had simply guessed. In a specific test scenario, the new method increased the quality of the spin information by a factor of nearly 1.7 compared to a baseline where the particle types were assigned randomly. This means that for the same number of events, the new method provides a much clearer picture of the quantum spin correlations. Furthermore, the system was able to reconstruct the spin correlation coefficient, a key number used to measure entanglement, with a value that matched the theoretical truth within the expected statistical limits.

This work demonstrates that the fully hadronic channel, long considered too messy for precision spin studies, can be tamed with the right tools. By successfully identifying the down-type quarks inside the jets, the researchers have unlocked a new window into the quantum behavior of the top quark. The method does not rely on any new physics or changes to the detector; it relies entirely on a smarter way of processing the data that is already being collected. The study confirms that with advanced machine learning, the most abundant decay channel of the top quark can now contribute to our understanding of quantum entanglement, offering a richer and more complete view of the subatomic world than was previously possible.

Technical Summary: Down-Type Jet Identification in Fully Hadronic $t\bar{t}$ Events

Problem Statement
Fully hadronic top-quark pair ( $t\bar{t}$ ) events offer the largest branching fraction ( $\simeq 45\%$ ) and provide down-type quarks as near-maximal spin analyzers ( $|\beta_d| \simeq 1$ ). However, these events are rarely utilized for spin-correlation and entanglement studies. The primary obstacle is the difficulty of reconstructing the specific down-type jets from the two hadronic $W$ decays amidst a large QCD multijet background and significant combinatorial ambiguity. While previous studies have addressed boosted top reconstruction or used constituent-level information in semileptonic channels, the resolved fully hadronic channel requires resolving the down/up ambiguity on both $W$ branches without a charged-lepton anchor, a task complicated by the exchange-equivalence of $W$ daughters in standard reconstruction benchmarks.

Methodology
The authors propose a two-stage machine-learning reconstruction framework designed to first assign jets to a legal top-pair candidate and then resolve the specific down/up identity of the $W$ -daughter jets.

Signal and Background Preparation:
- Signal: Simulated $t\bar{t}$ events at $\sqrt{s}=13$ TeV using MadGraph5_aMC@NLO, MadSpin, PYTHIA, and DELPHES. Events are preselected with $p_T > 30$ GeV, $|\eta| < 2.4$ , and specific $H_T$ and $b$ -tagging requirements. Truth labels are assigned via geometric matching to generator-level partons.
- Background: Includes QCD multijets (normalized to ATLAS all-hadronic measurements) and unmatched $t\bar{t}$ events (where jets cannot be fully matched to partons).
- Inputs: The model utilizes 14-dimensional jet-level features (including kinematics, $b$ -tagging, and shape variables) and up to 40 particle-flow constituents per jet (10-dimensional vectors including relative coordinates, energy, charge, and particle type).
Network Architecture (Hybrid GNN+Transformer+Diffusion):
- Step-1 (ST1) Classifier: A Graph Neural Network (GNN) combined with a Transformer processes all reconstructed jets in an event. It outputs logits for five coarse labels: $\{other, b_1, W_1, b_2, W_2\}$ . A legal six-jet assignment is selected by maximizing an assignment score that combines the neural network logits with a mass prior term (penalizing deviations from $m_W$ and $m_t$ ).
- Step-2 (ST2) Classifier: Takes the six jets selected by ST1 (ordered as $b_1, q_{1a}, q_{1b}, b_2, q_{2a}, q_{2b}$ ) and resolves the down/up ambiguity. It enumerates four possible hypotheses for the down-type jets in the two $W$ branches.
- Auxiliary Diffusion Head: ST2 includes a conditional diffusion denoiser that shares the event context representation. This head is trained to predict noise injected into the assignment hypothesis tokens. Crucially, the authors introduce a margin ranking objective for this diffusion head, requiring the diffusion score of the true assignment to exceed competing hypotheses by a margin.
Training and Loss Functions:
- ST1 is trained with a cross-entropy loss invariant to the exchange of the two top branches.
- ST2 uses a standard cross-entropy loss for the assignment scorer combined with the diffusion margin loss. The authors found that standard denoising objectives alone failed to make the diffusion score discriminative for ranking; the margin objective was essential to lift the ranking accuracy above random baseline.

Key Contributions

Two-Stage Reconstruction: A novel pipeline that separates the coarse jet-to-parton assignment from the specific down/up flavor identification, utilizing constituent-level information to break the $W$ -daughter symmetry.
Diffusion for Ranking: The demonstration that a conditional diffusion head, when trained with a margin ranking objective rather than a standard denoising objective, can effectively rank discrete assignment hypotheses. This provides an auxiliary signal for training and additional cut variables.
Spin-Analysis Preservation: A rigorous evaluation showing that down/up identification in the fully hadronic channel can retain spin-correlation information despite combinatorial ambiguities and background contamination.

Results

Reconstruction Quality: At a representative working point ( $Q^{avg}_{ST1} \ge -0.0889$ ), the method achieves a signal efficiency of $\epsilon \simeq 0.0845$ with a calibrated background-to-signal ratio of $N_{bkg}/N_{sig} = 0.631$ .
Spin-Analysis Performance: Compared to a benchmark using random down/up labels (with identical ST1 selection), the learned ST2 classifier improves the effective spin-analysis factor ( $\eta_\kappa$ ) from 0.152 to 0.255 (a factor of 1.68) and the purity-weighted six-parton exact fraction from 0.139 to 0.218 (a factor of 1.57).
Spin Correlation Coefficient ( $D$ ): The reconstructed spin-correlation coefficient $D$ is statistically compatible with the truth-level value within resampling uncertainties. Specifically, for signal plus unmatched events, the ML-reconstructed $D$ is closer to the truth value than the "soft jet" prescription (which selects the lower-energy light jet).
Diffusion Objective Analysis: The study confirms that standard denoising objectives leave the diffusion assignment score at the random baseline, whereas the margin ranking objective makes it discriminative. The two objectives are largely decoupled.

Significance
The paper claims that down/up identification in the all-hadronic channel is feasible and capable of retaining useful spin-correlation information. By successfully resolving the down-type quark direction in fully hadronic $t\bar{t}$ events, this work enables the use of the channel with the largest branching fraction for quantum tomography, Bell inequality tests, and entanglement studies, which have previously been limited to dilepton or lepton+jets channels due to reconstruction difficulties. The results suggest that the fully hadronic channel, often disfavoured for spin studies, can be made experimentally relevant through advanced machine-learning reconstruction techniques.

More like this