Long-lived particles and dark sectors

Many dark sector models predict particles that travel a measurable distance before decaying. I led a CMS search for such particles produced with a Z boson and decaying to displaced jets in the tracker, which extended the sensitivity to light particles, down to about 15 GeV. I helped develop a method that uses the endcap muon detectors as a sampling calorimeter; the first search using it set the most stringent limits at the time for proper decay lengths above 6 to 40 m, depending on the mass. Most recently I led a search in the 2018 B-parking dataset, about ten billion b-hadron decays, which set the most stringent limits to date on B → KΦ for long-lived Φ of 0.3 to 3 GeV.

Real-time machine learning

The CMS Level-1 trigger has a few microseconds to decide which of 40 million bunch crossings per second to keep, so any model running there has to fit in FPGA firmware. I set up an FPGA test stand at UC San Diego and led a team of students developing a convolutional neural network that tags long-lived particle decays in the tracker and calorimeters; we also used deep neural networks to improve the real-time calculation of missing transverse energy, with models compiled to firmware using hls4ml. I am now co-PI of AIDA-Scout, a DOE-funded project on AI for real-time anomaly detection in the CMS Level-1 scouting system.

Machine learning for physics analysis

In offline analysis I supervise the development of neural-network and boosted-decision-tree taggers that identify long-lived particles decaying in the CMS muon system and measure their energy, and my students have extended this work to graph convolutional networks. With collaborators I have also worked on graph neural network methods more broadly: graph autoencoders for anomaly detection on jets, trained with a learned, differentiable energy mover's distance, and layerwise relevance propagation to explain a graph neural network that performs particle-flow reconstruction. I also studied graph neural network jet taggers for Higgs boson pair production at the FCC-hh.

AI infrastructure for science

At the San Diego Supercomputer Center I operate large-scale, Kubernetes-based analysis infrastructure on the National Research Platform, including self-hosted open-weight large language models, for a broad community of scientists with ML- and data-intensive workflows. I also teach hands-on training on the platform, covering Kubernetes, hosted LLMs, retrieval-augmented generation and AI agents; the materials are at training.nrp-nautilus.io and much of the code is on my NRP GitLab. Through the FAIR4HEP project I helped develop guidance for making AI models and datasets in high energy physics findable, accessible, interoperable and reusable.

Eighteen students have worked with me: graduate, undergraduate, postbaccalaureate and high school. They have built neural-network, graph-network and boosted-decision-tree taggers, worked on analyses and produced results that appear in CMS publications. Two graduate students and an undergraduate worked on the B-parking search under my supervision. From 2023 to 2025 I chaired the US CMS Mentorship Committee, and I serve on the Equity and Career Committee of A3D3, the NSF institute for Accelerated AI Algorithms for Data-Driven Discovery.

I teach machine learning, triggers and jet substructure in the Hands-On Advanced Tutorial Sessions (HATS) at the Fermilab LPC, have taught at the CMS Data Analysis School, and co-organized an hls4ml tutorial at the 2022 Snowmass summer meeting in Seattle. For the National Research Platform I teach researchers, students and faculty to run AI workloads, including hosted LLMs and AI agents, in seven hands-on trainings and webinars in 2026. I am always glad to hear from students looking for a way into this field.