Speaker
Description
The design of modern accelerator and RF components demands high-fidelity, multi-physics modeling at scale, motivating the tight integration of AI/ML methods with high-performance computing (HPC). ACE3P, a comprehensive suite of conformal, high-order, parallel finite-element codes developed at SLAC, provides coupled electromagnetic, thermal, and mechanical analysis for a broad range of accelerator components—from crab cavities and RF guns to distributed-coupling linacs and multipacting studies. Built on curved high-order finite elements and deployed on massively parallel platforms such as NERSC's Perlmutter, ACE3P leverages hybrid MPI+OpenMP parallelization and GPU offloading to achieve the fidelity and throughput required for realistic virtual prototyping.
Building on this foundation, we present AI/ML-driven workflows that accelerate the co-design of accelerator components. Multi-fidelity Bayesian optimization, implemented through the LUME-ACE3P-Xopt framework, substantially reduces design cycles for RF cavities and components by intelligently balancing fast low-fidelity and accurate high-fidelity simulations, outperforming brute-force parameter sweeps. Surrogate Gaussian-process models further enable the solution of inverse problems, such as identifying dark-current field-emission origins that would otherwise require daunting trial-and-error simulation campaigns. Looking ahead, we discuss the integration of RF and radiation simulation (ACE3P + Geant4) and the development of agentic workflows to automate multi-physics simulation pipelines end-to-end. Together, these capabilities illustrate how AI/ML integrated with HPC-based modeling can serve as a cornerstone for a dedicated AI/ML accelerator test facility, informing requirements for computing infrastructure, data handling, and configurable simulation-driven design.
