Using the top quark to look for new physics

01 · Top quark physics

Four top quarks in one collision

• four-top production
• top-Higgs coupling
• machine learning
• statistical combination
• LHC Run 2

At the LHC, protons collide at the highest energies ever reached in a laboratory, and the CMS detector records what comes out. The top quark is roughly as heavy as an atom of gold and is the elementary particle most strongly coupled to the Higgs boson, which makes it a natural place to look for what lies beyond the Standard Model.

It also decays in less than a trillionth of a trillionth of a second, before it has time to bind with other quarks. Its decay products therefore carry the top quark’s own properties with them, which makes it a clean probe of the Higgs boson and of possible new particles.

Producing four top quarks in a single collision is extremely rare, and how often it happens depends on the top quark’s coupling to the Higgs boson and on any new heavy particles that could take part. At Brown University I searched for this process in the single-lepton final state (one electron or muon plus many jets).

I worked on the machine-learning tools that pick these events out of far more common backgrounds, and on the statistical combination with other final states using the complete LHC Run 2 data set (recorded from 2015 to 2018). I also prepared and optimized the analysis for Run 3.

• I was a major contributor to the CMS paper reporting evidence for four-top-quark production (Physics Letters B, 2023), which includes this search and the combination.
• A review of four-top physics at the LHC, written with F. Blekman, F. Déliot and V. Dutta (Universe, 2022).
• A talk and proceedings at LHCP 2021 reporting ATLAS and CMS results on κ_t, the strength of the top quark’s coupling to the Higgs boson relative to the Standard Model prediction, from measurements of top quark-antiquark pairs (tt̄) and of four top quarks.

At Alabama, two graduate students in my group carry this program forward: Guillermo Fidalgo in the single-lepton final state, and Nathan Nguyen in the opposite-sign dilepton final state (two electrons or muons of opposite electric charge).

04 · Artificial intelligence

AI for particle physics

• end-to-end deep learning
• convolutional networks
• graph neural networks
• CMS Open Data
• data quality

Collider analyses usually compress each collision into a short list of reconstructed objects, such as jets, electrons and muons, before any machine learning is applied. With a team at Google Research, we explored a different route: end-to-end deep learning, in which convolutional and graph neural networks learn directly from low-level detector data, the hits in the tracker and the energy deposits in the calorimeters you see in the background.

We used it to identify particles and whole collision events in both regular and boosted topologies, and extended it to top-quark mass regression. At Alabama I continue to work on improving particle reconstruction and event identification with AI.

So that others can reproduce and build on this work, I produced CMS Open Data sets for machine learning with low-level detector inputs, including tracker hits. As manager of Data Quality Monitoring and Data Certification in CMS, I worked on machine-learning-based data certification to significantly reduce the human input needed to decide which data are good for physics.

• End-to-end classification of boosted top-quark jets, and of quark versus gluon jets, using CMS Open Data (talks include a plenary at BOOST 2019).
• Co-author of the Machine Learning in High Energy Physics Community White Paper.
• Co-administrator of the Machine Learning for Science (ML4SCI) organization at Google Summer of Code since 2021, and of HumanAI since 2024. In 2021 I coordinated 19 ML4SCI projects, including quantum generative models and quantum convolutional neural networks for physics data analysis.