Machine learning in top quark physics at ATLAS and CMS

This note presents an overview of current and potential future applications of machine-learning-based techniques in the study of the top quark. The research community has developed a diverse set of ideas and tools, including algorithms for the efficient reconstruction of recorded collision events and innovative methods for statistical inference. Recent applications of some techniques by the ATLAS and CMS collaborations are also highlighted.

💡 Research Summary

This paper provides a comprehensive overview of how machine‑learning (ML) techniques are currently employed and future‑oriented within top‑quark physics at the ATLAS and CMS experiments. The introduction emphasizes that ML has been a driving force for over a decade, enabling key milestones such as single‑top discovery, b‑jet identification improvements, and the observation of four‑top production.

The reconstruction section is split into two tasks. First, the inference of the neutrino direction in semileptonic top decays is tackled by the ν‑Flow method, which uses a conditional normalizing‑flow neural network to map the true neutrino vector onto a three‑dimensional Gaussian. Sampling from this distribution yields a likelihood for possible neutrino directions, outperforming traditional feed‑forward regressors and simple W‑mass constraints. Second, the assignment of the remaining decay products to the correct top quark is addressed by several approaches. SPANET employs a transformer‑based architecture with more than 10 million parameters and auxiliary targets (neutrino regression and signal‑background discrimination) to achieve state‑of‑the‑art performance. The HYPER method represents decay products as hypergraphs, allowing edges to connect multiple nodes; despite using only 345 k parameters, its performance rivals SPANET.

In the analysis‑strategy part, the paper discusses the difficulty of estimating QCD multijet backgrounds. The classic ABCD matrix method is automated by the DISC technique, which trains a classifier while adding a penalty term that suppresses correlations between the classifier score and the two independent observables, thereby preserving the ABCD assumption. A recent CMS all‑hadronic four‑top search employed an autoregressive normalizing flow to map events from a background‑enriched region into the signal region, providing a data‑driven background estimate.

The statistical‑inference section moves beyond binned likelihoods. Likelihood‑free (simulation‑based) inference uses the classifier output s as a direct test statistic, exploiting the relation (L_{1}/L_{0}=s/(1-s)). Tools such as INFERNO and SALLY implement this idea while propagating systematic uncertainties. Unfolding, an intrinsically ill‑posed inverse problem, is tackled by OMNIFOLD, which iteratively trains a classifier to reweight simulated events toward data, allowing unbinned, multidimensional unfolding. The method has been demonstrated by ATLAS (average jet mass vs. jet pT in Drell‑Yan) and CMS (charged constituent multiplicity in minimum‑bias events).

Looking toward the High‑Luminosity LHC, the paper highlights the growing computational burden of large data sets. CMS introduced the DCTR (Neural reweighting) technique to emulate parameter variations—such as the POWHEG h‑damp parameter—or to upgrade NLO samples to NNLO accuracy without generating new Monte‑Carlo samples. This approach promises substantial sustainability gains by reducing the need for full detector simulations.

The conclusion reiterates that ML has become indispensable across top‑quark reconstruction, background estimation, and statistical inference, and that continued development of efficient, uncertainty‑aware algorithms will be crucial for the precision era of the HL‑LHC. The reference list provides a solid bibliography covering b‑jet tagging, ν‑Flow, SPANET, HYPER, DISC, INFERNO, SALLY, OMNIFOLD, and DCTR.