Machine learning in top quark physics at ATLAS and CMS

This paper provides an overview of current and future machine learning applications in top quark physics, highlighting diverse algorithms for event reconstruction and statistical inference used by the ATLAS and CMS collaborations.

The Gist

Imagine the Large Hadron Collider (LHC) as the world's most powerful, high-speed camera, snapping pictures of particles smashing into each other at nearly the speed of light. Among the billions of particles created, the "top quark" is a superstar—it's the heaviest and most unstable, decaying almost instantly into other particles. The paper you provided is a report card on how scientists at the ATLAS and CMS experiments are using Machine Learning (ML)—a type of computer intelligence—to make sense of this chaotic cosmic debris.

Here is a breakdown of their work using everyday analogies:

1. The Detective Work: Finding the Invisible

When a top quark decays, it sometimes produces a neutrino. Think of a neutrino as a ghost: it passes through the detector without leaving a single trace, making it invisible. However, physicists know it must be there because energy and momentum have to balance out.

• The Old Way: Trying to guess where the ghost went by drawing straight lines or using simple math rules.
• The New ML Way: The paper highlights tools like ν-FLOW and SPANET. Imagine these as super-detectives that have studied millions of crime scenes. Instead of just guessing, they look at the "footprints" left by the visible particles and use a complex internal map (a neural network) to predict exactly where the invisible ghost is most likely to be.
  • ν-FLOW is like a detective who draws a cloud of possible locations for the ghost, showing you the most probable spots.
  • SPANET is like a master organizer that not only finds the ghost but also sorts all the other scattered debris (jets and leptons) to figure out which piece belongs to which original top quark. It's so good it uses over 10 million "brain cells" (parameters) to do this.
  • HYPER is a newer, lighter detective. It uses a clever trick called "hypergraphs" (where one connection can link many things at once) to solve the same puzzle with far fewer resources, yet just as accurately.

2. Sorting the Noise: The "ABCD" Strategy

In these experiments, the signal (top quarks) is often hidden in a mountain of "noise" (background events caused by other particle interactions). It's like trying to find a specific type of rare coin in a pile of millions of regular coins and trash.

• The Challenge: Some of the "trash" (background) looks exactly like the "coins" (signal), making it hard to count them accurately.
• The Solution: The paper discusses the DISCO method. Imagine you have two different sorting machines. Usually, they might get confused and mix things up. DISCO trains a computer to build two sorting criteria that are completely independent of each other (like sorting by color and then by weight, where one doesn't affect the other). This allows the scientists to use data from "safe" areas to accurately predict how much noise is in the "dangerous" areas where the signal is hiding.
• Another Trick: For a specific search involving four top quarks crashing together, the CMS team used a tool that acts like a time machine. It takes events from a "background-heavy" zone and mathematically transforms them to look like they came from the "signal" zone, helping them understand the background better without needing to simulate it from scratch.

3. The Final Verdict: Better Statistics

Once the data is sorted, scientists need to decide: "Is this a real discovery or just a fluke?"

• Likelihood-Free Inference: Traditionally, this is like calculating odds using a rigid formula. The new ML tools (like INFERNO and SALLY) act more like a smart judge. Instead of just crunching numbers, they look at the "score" a computer gives to an event and use that score directly to decide if a hypothesis is true or false. It's a faster, more flexible way to weigh the evidence.
• Unfolding the Truth: Sometimes, the detector blurs the picture, making a sharp line look fuzzy. "Unfolding" is the process of sharpening that image to see the true shape.
  • The OMNIFOLD method is like a smart photo editor. It compares the blurry photo (the data) with a perfect reference photo (the simulation). It learns the differences and then "reweights" the data, effectively sharpening the image to match reality.
  • The paper notes this allows them to measure things in multiple dimensions at once, like seeing how the "weight" of a jet changes as its "speed" changes, all without losing detail.

4. The Future: The High-Luminosity LHC

The LHC is about to enter a "High-Luminosity" phase, meaning it will produce massive amounts of data—far more than computers can currently handle by running slow, traditional simulations for every single possibility.

• The Problem: Simulating every possible scenario is like trying to paint a masterpiece by hand for every single frame of a movie. It takes too long and uses too much energy.
• The ML Solution (DCTR): The CMS collaboration introduced a method called DCTR. Think of this as a smart filter or a digital chameleon.
  • Instead of generating a brand new simulation for every tiny change in physics parameters, they take one existing simulation and use ML to "reweight" it.
  • Analogy: If you have a photo of a sunny day, DCTR can digitally adjust the lighting to make it look like a cloudy day or a sunset without taking a new photo.
  • The paper shows this works for adjusting complex physics settings (like the energy of radiation) and even for upgrading the accuracy of the math (turning a "good" approximation into a "perfect" one). This saves massive amounts of computing power and time.

Summary

In short, this paper explains that Machine Learning has moved from being a "nice-to-have" tool to the engine driving top quark research. It helps physicists:

1. Find the invisible (neutrinos).
2. Sort the noise from the signal efficiently.
3. Make better statistical decisions about what they've found.
4. Prepare for the future by making simulations faster and smarter, ensuring they can handle the data deluge of the next generation of the LHC.

The authors conclude that these tools are not just helping them understand the top quark today, but are essential for the high-precision discoveries they hope to make tomorrow.

Technical Summary

Technical Summary: Machine Learning in Top Quark Physics at ATLAS and CMS

Problem Statement

The study of the top quark at the Large Hadron Collider (LHC) faces significant challenges in event reconstruction, background estimation, and statistical inference. Specifically, the field requires:

1. Efficient Reconstruction: Determining the kinematics of undetected neutrinos in leptonic top decays ($t \to b\ell\nu$) and correctly associating decay products (leptons and jets) to specific top quarks in complex events.
2. Background Modeling: Accurately estimating background rates for multijet events, particularly those arising from pure QCD interactions, which are difficult to predict via standard simulation.
3. Statistical Inference: Moving beyond traditional binned likelihood approaches to improve the extraction of physical parameters and the unfolding of differential cross-sections.
4. Future Scalability: Addressing the increased computing demands of the upcoming High-Luminosity LHC (HL-LHC) by reducing reliance on computationally expensive simulated samples and detector simulations.

Methodology

The paper reviews a diverse set of machine learning (ML) algorithms and frameworks currently employed or proposed by the ATLAS and CMS collaborations:

• Neutrino Inference:

  • $\nu$-FLOW: Utilizes a normalizing flow neural network conditioned on reconstructed event observables. It maps the true neutrino direction vector to a 3D normal distribution, allowing for the inference of likelihoods for possible neutrino directions by sampling, rather than simple regression.
  • SPANET: Employs a neural network transformer architecture (over 10M parameters) to assign all top decay products to reconstructed particles. It incorporates auxiliary targets, such as neutrino direction regression and signal/background discrimination.
  • HYPER: A novel approach representing decay products as hypergraphs (generalizing graph NNs where edges connect more than two nodes). It achieves performance comparable to SPANET with significantly fewer parameters (345k).

• Analysis Strategies:

  • DISCO: Introduces an NN classifier to construct observables that are uncorrelated and effectively separate signal from background. This is achieved via a penalty term during training to suppress distance correlations between classifier scores or between a score and an auxiliary observable.
  • Auto-regressive Normalizing Flows: Used in CMS analyses to transform data events from background-enriched regions into signal regions for all-hadronic four-top-quark searches.

• Statistical Inference and Unfolding:

  • Likelihood-free Inference: Tools like INFERNO and SALLY use the output score ($s$) of a classifier as a test statistic, exploiting the relationship $H_1/H_0 = s/(1-s)$ for hypothesis testing while accounting for systematic uncertainties.
  • OMNIFOLD: Facilitates unbinned, multidimensional unfolding of differential cross-sections. It uses an iterative procedure where a classifier learns differences between simulation and data, subsequently reweighting the simulated sample to match data distributions. The number of iterations controls regularization.

• HL-LHC Optimization (Reweighting):

  • DCTR (Deep Classifier for Reweighting): A method used by CMS to reweight simulated samples to emulate parameter shifts (e.g., the hdamp parameter in POWHEG) or to achieve higher-order accuracy (e.g., reweighting NLO samples to match NNLO predictions). This aims to replace the generation of dedicated samples for systematic variations.

Key Results

The paper highlights several successful applications and performance metrics:

• Reconstruction: The $\nu$-FLOW approach demonstrates superior performance in inferring neutrino pseudorapidity compared to feed-forward NN regression or W boson mass constraints. HYPER achieves SPANET-level performance with a fraction of the parameters.
• Background Estimation: The DISCO method successfully constructs uncorrelated observables for signal/background separation in multijet environments.
• Unfolding: OMNIFOLD has been successfully demonstrated by ATLAS and CMS for unfolding Drell-Yan and minimum bias events, respectively. Notably, its unbinned nature allows for the unfolding of novel quantities, such as the average jet mass as a function of jet $p_T$.
• Reweighting: The DCTR method shows good agreement when reweighting samples to emulate hdamp variations and when upgrading NLO samples to NNLO accuracy, suggesting a viable path to reducing computational costs.

Significance and Claims

The paper positions machine learning as a "driving force" in top quark physics for over a decade, citing its critical role in milestones ranging from single-top production at the Tevatron to the recent observation of four-top-quark events at the LHC.

The authors claim that:

1. Current Impact: ML algorithms are essential for efficient event reconstruction and innovative statistical inference, directly enabling the observation of rare top quark processes.
2. Future Outlook: Novel ML-based approaches in reconstruction, background estimation, and statistical inference are "setting the stage" for the high-precision era of the HL-LHC.
3. Computational Sustainability: Techniques like DCTR offer a pathway to improve sustainability by potentially skipping the computational needs of classical detector simulations and dedicated sample generation for systematic uncertainties.

The paper concludes that while no ML algorithm can overcome the inherent ill-posed nature of unfolding problems (requiring regularization), the integration of these tools provides valuable contributions to the field and prepares the community for future data challenges.