15–18 Dec 2026
Duke University
America/New_York timezone
Artificial Intelligence for the Electron Ion Collider (2026)

Harness Engineering for AI-Assisted Development of EIC Beam-Beam Simulations

Not scheduled
15m
LSRC B101 Love Auditorium (Duke University)

LSRC B101 Love Auditorium

Duke University

Durham, NC 27710
AI/ML for Accelerators AI/ML for Accelerators

Speaker

Derong Xu (Brookhaven National Laboratory)

Description

Beam-beam interactions are a major factor limiting luminosity in the Electron-Ion Collider (EIC), motivating reliable and efficient simulations. AI coding agents offer new opportunities for developing high-performance tools to simulate these complex, nonlinear interactions, but their effective use requires explicit physical assumptions, consistent numerical conventions, and systematic verification. We share our experience with harness engineering for AI-assisted development of EIC beam-beam simulation tool. Here, the harness refers to the specifications, scientific context, software interfaces, and executable checks that guide agent contributions and support human review.

We illustrate these experiences through the development of a Julia framework supporting GPU-accelerated weak-strong and self-consistent strong-strong beam-beam simulations. Architectural specifications and knowledge notes connect implementation choices to the underlying physics, while modular interfaces support alternative field solvers and CPU/GPU execution. For an EIC strong-strong simulation using the production particle-in-cell solver, runtime decreased from 35 hours with our previous C++ MPI implementation to 2.5 hours with the new implementation on a NERSC A100 GPU, with results showing good agreement with BeamBeam3D. Verification combines software consistency checks with analytical limits and comparisons against established beam-beam calculations. Together, these elements help evaluate agent-generated changes against computational requirements and physical expectations.

We discuss practical lessons from human-AI collaborative development, including the organization of scientific context, the role of executable checks in identifying implementation errors, and the distinction between software correctness and physical validation. EIC beam-beam applications and computational benchmarks provide concrete examples of the resulting capabilities. We also discuss current limitations and possible extensions toward agent-guided simulation studies with explicit validation and human oversight.

Author

Derong Xu (Brookhaven National Laboratory)

Co-author

Yi-Kai Kan (Brookhaven National Laboratory)

Presentation materials

There are no materials yet.