IPAC'26 - the 17th International Particle Accelerator Conference

SciBmad: A differentiable, GPU-parallelized software library for particle accelerator design, nonlinear analysis, and machine learning

THP5325

21 May 2026, 16:00

2h

C.I.D

Poster Presentation MC5.D02: Nonlinear Single Particle Dynamics Resonances, Tracking, Higher Order, Dynamic Aperture, Code Developments Poster session

Speaker

Joseph Devlin (Cornell University (CLASSE))

Description

We present SciBmad, a new accelerator physics software library developed to meet the needs of all accelerator design, analysis, and virtual modeling. SciBmad consists of a set of modular packages featuring fully differentiable and CPU/GPU-parallelized symplectic integrators, spin tracking and radiation, flexible and differentiable lattice definitions, nonlinear normal form analysis tools with Lie algebraic methods, batch parameter parallelization, and more. It is fully usable in Julia and easily callable from Python, making it easy to integrate it with external optimizers and machine learning frameworks. All releases of SciBmad undergo rigorous, automated testing with maximal code coverage, and all integrators are validated against PTC. With a growing list of features and contributors, SciBmad aims to be a powerful tool for any accelerator physics application.

Funding Agency

U.S. Department of Energ Contracts No. DE-SC0012704 (Brookhaven Science Associates, LLC) and No. 89233218CNA000001 (Office of Science, ARDAP), and NERSC under award HEP-ERCAP0034811.