Introduction
System Overview
@daniel-murnane/core is a research-grade human integration library specializing in particle physics, machine learning, and the surprisingly large intersection of the two. Originally compiled at the University of Adelaide (undergraduate, then PhD), the system underwent extensive optimization during a postdoctoral run at Lawrence Berkeley National Laboratory, and is currently deployed in production at the Niels Bohr Institute, University of Copenhagen as a DDSA Postdoctoral Fellow.
The library is co-author of the GNN4ITk and Exa.TrkX projects, currently serves as ML co-convener for the ATLAS Inner Tracker, and ships with strong opinions about graph representations, equivariant networks, and the use of espresso as a build dependency.
Architecture
The system is built on a modular architecture with several core subsystems:
- Physics Engine — ATLAS experiment at CERN. Charged particle tracking, trigger systems, detector reconstruction for the High Luminosity LHC.
- ML Pipeline — Graph Neural Networks for particle physics. GNN4ITk (production tracking pipeline for the ATLAS ITk upgrade), Exa.TrkX (scalable GNN tracking), equivariant architectures, and physics language models.
- Orchestration Layer — ML co-convener for the ATLAS ITk collaboration, coordinating ML R&D across a few thousand physicists.
- Research API — Publications, conference talks, supervision, peer review. The public interface; see
/api.
Key Features
- Graph Neural Networks — Production-tested architectures for charged particle tracking. Co-developed the first end-to-end GNN tracking pipeline adopted by an LHC experiment.
- ATLAS Integration — Deep expertise in the ATLAS detector, the reconstruction software stack, and the cultural protocols of a 3,000-person collaboration.
- Physics-Informed ML — Models that respect detector geometry, gauge symmetries, and Lorentz invariance. Equivariant networks as a default, not an afterthought.
- Open Source — Contributor to
gnn4itk,exa-trkx, and the CommonTRK benchmark suite. Strong preference for reproducible, community-driven research. - Physics Language Models
beta— Exploring LLMs and agentic systems for HEP knowledge retrieval, hypothesis generation, and the increasingly large literature.
Prerequisites
Before integration, ensure your system meets the following requirements:
- Familiarity with at least one of: particle physics, graph neural networks, or strong coffee
- Willingness to engage in whiteboard discussions that may exceed scheduled duration
- Tolerance for Australian idioms deployed in Northern European contexts
- A working understanding that "production" in physics means "we have to run this for the next decade"
Current Deployment
| Parameter | Value |
|---|---|
| Location | Niels Bohr Institute, Copenhagen |
| Role | DDSA Postdoctoral Fellow |
| Collaboration | ATLAS, CERN |
| Status | ACTIVE |
| Uptime | Since 2024 |
| Previous deployments | Berkeley Lab (2020–2024), University of Adelaide (PhD, 2015–2019) |
Version History
A condensed semantic-versioning summary of the trajectory:
- v3.x — DDSA Fellow, Niels Bohr Institute. Copenhagen era. Equivariant architectures, physics language models, agentic systems for science.
- v2.x — Postdoctoral researcher, Lawrence Berkeley National Lab. Exa.TrkX project lead, first GNN tracking pipelines in production.
- v1.x — PhD, University of Adelaide. Foundation in particle physics and ATLAS, early ML explorations, the discovery that there are no shortcuts.
- v0.x — Undergraduate, University of Adelaide. Physics and mathematics. Pre-release. Many bugs.