Chen-Hsun (Jay) Chan
Machine Learning Postdoctoral Research Fellow at Lawrence Berkeley National Laboratory
50-5047
1 Cyclotron Road
Berkeley, CA 94720
I’m a physicist and machine learning researcher in the Scientific Data Division at Berkeley Lab. I develop models and inference systems for scientific workloads, with a focus on graph neural networks, GPU performance, and reinforcement learning.
For ATLAS, I develop GNN4ITk to reconstruct charged-particle tracks from detector hits. I introduced Double Metric Learning for sparse tracking graphs. I proposed and developed EggNet, an evolving graph attention network I presented at the ICML 2024 AI for Science Workshop. I presented D-WALK and Junction Resolution at CHEP 2026. I’m now working on Tracking-as-a-Service to benchmark and improve GPU inference for tracking.
I also develop reinforcement learning for on-chip packet routing in NIFTY and co-developed an AI agent system for high-energy physics analysis. My earlier work includes detector unfolding, generative hadronization, and ATLAS Higgs boson and dark matter analyses. I earned my PhD in Physics from the University of Wisconsin–Madison in 2023.
Research projects · Publications · CV and résumé
Outside the lab, I enjoy hiking, skiing, and playing piano and drums.
news
| May 25, 2026 | I presented D-WALK and Junction Resolution for GNN-based track finding at CHEP 2026. |
|---|---|
| May 15, 2026 | My Double Metric Learning paper on building directed tracking graphs for ATLAS ITk is available in the Connecting the Dots proceedings. |
| Jul 26, 2024 | I presented EggNet, an evolving graph attention network for particle tracking, at the ICML 2024 AI for Science Workshop in Vienna. |
| Dec 15, 2023 | The new paper of HadML (integrating particle flavors) is now on arXiv (2312.08453)! |
| Jun 1, 2023 | The 2HDM+a combination paper is now on arXiv (2306.00641)! |
selected publications
-
Double Metric Learning for Building Directed Graphs with Chain Connections for the ATLAS ITk DetectorIn Proceedings of Connecting the Dots 2025, 2026