Particle Tracking with GNNs
Particle tracking is the connecting-the-dots process of linking tracker hits to form particle trajectories; the resulting hit cluster is referred to as a track. Tracker data is naturally represented as a graph by assigning hits to nodes and drawing edges that represent hypothesized particle trajectories. Edge-classifying graph neural networks (GNNs) are trained to predict the probabilities, or edge weights, that each edge links two hits generated by the same particle. I build edge-classifying GNN pipelines that (1)Â build graphs from raw data, (2) classify edges, and (3) cluster hits based on edge classification scores. One such pipeline, based on the Interaction Network architecture, is implemented below. I am also studying computer vision approaches to particle tracking, for example by predicting bounding boxes around tracks.
Displaced Taus
Displaced taus are a challenging but rewarding place to search for Beyond the Standard Model (BSM) physics. Many well-motivated BSM scenarios give rise to long lived particles (LLPs) decaying to taus; for example, X→τ + Ɇt is a signature of gauge-mediated SUSY breaking (where X is a tau slepton). Taus themselves decay hadronically with a branching fraction of 65%, otherwise decaying leptonically. To date there are no dedicated searches for displaced τ, representing an unaddressed gap in the LHC physics coverage. In order to study LLPs decaying to taus, it is therefore critical to develop strategies to reconstruct, identify, and, in the HL-LHC era, trigger on displaced τ. In anticipation of a dedicated displaced τ analysis, I am studying graph neural network (GNN) algorithms for displaced $\tau_h$ reconstruction from underlying tracker and calorimeter hits.
A→Zh Analysis
The heavy pseudoscalar Higgs boson A is to date unobserved, but appears in many well-motivated extensions of the SM such as the minimal supersymmetric SM. This analysis targets A decaying to a Z boson and a Standard Model (SM) Higgs boson h, where the Z decays to two light leptons and the h decays to two taus. Summarily, this process is A→Zh→llττ, or AZh for short. I am leading a search for AZh with llττ final states through state-of-the-art computational tools like Coffea and Awkward Array. The analysis framework is available on my Git page, linked below.
​
(Left, Top) The Feynman diagram for A production via gluon-gluon fusion. (Left, Bottom) The analysis targets the corrected and constrained (via SVfit) mass of the llττ final state; this plot was published in HIG-18-023.
CMSML Documentation and Tutorials
The CMS Machine Learning (CMSML) Knowledge subgroup is building a set of tutorial and documentation pages for various ML tools used in the CMS software (CMSSW). I have contributed the PyTorch and PyTorch Geometric documentation pages, which can be found in the tutorials/inference subdirectory.
