Projects
A selection of my research projects.
Cross-Domain Transfer with Particle Physics Foundation Models
Transferring jet foundation models to MINERvA neutrino interactions
Foundation models trained on collider jet data learn representations of jets that turn out to be useful outside their original domain. In this project, we study how well representations learned from jet physics transfer to neutrino interaction data from the MINERvA experiment, probing how much structure these models share across very different detector and physics regimes. Using foundation models trained on high-energy jets at the medium-energy MINERvA experiment leads to faster training and better performance achieved with the same amount of data.
Explore a sample of MINERvA events used in the study in the interactive event viewer, and check out the paper for more info.
🔭 Explore the event viewer 📄 Read the paper 💻 View the code
HitPF
End-to-end hit-level particle flow reconstruction for future colliders
HitPF is an end-to-end, Geometric Algebra Transformer-based particle flow reconstruction pipeline for that works directly on detector hits and tracks. By learning reconstruction directly from raw hits rather than hand-engineered intermediate objects, the approach aims to push the precision of event reconstruction beyond what classical particle flow algorithms achieve. By decoupling reconstruction performance from detector-specific tuning, this framework enables rapid iteration during the detector design phase of future collider experiments.
Run the algorithm yourself on any event on Hugging Face, and check out the paper for more info.
🤗 Try the live demo 📄 Read the paper 💻 View the code
Semi-Visible Jet Clustering
Learning IRC-safe jet clustering with Geometric Algebra Transformers
We use Lorentz-equivariant Geometric Algebra Transformers to learn jet clustering algorithms that remain infrared- and collinear-safe, with an eye towards reconstructing semi-visible jets, where part of the jet’s energy escapes into invisible or weakly-interacting particles. Learning the clustering directly lets the model adapt to these harder topologies. The models outperform anti-kt, the usual jet clustering algorithm used at the Large Hadron Collider (LHC) experiments.