SPARC digital twin to use DOE supercomputers for AI pulse planning

The digital twin is intended to support pulse planning in the SPARC control room while the tokamak is still being assembled in Devens, Massachusetts

(Image courtesy of Commonwealth Fusion Systems)

Commonwealth Fusion Systems (CFS) will lead a SPARC digital twin project in the second phase of the US Department of Energy’s Genesis Mission, the only private company selected to head a Phase II award this month. The platform will draw on DOE supercomputers and decades of fusion data to support operations of SPARC, the tokamak CFS is assembling in Devens, Massachusetts, and it is the only project in this round focused on fusion.

A SPARC digital twin built on fusion data

The workstream covers software infrastructure to connect with DOE computing resources, the training datasets and AI surrogate models a digital twin requires, reinforcement learning and agentic AI models for an AI-based SPARC operations co-pilot, and data visualisation pipelines that accelerate operations. It builds on digital twin visualisation work CFS began earlier this year with NVIDIA and Siemens, and DOE announced the Phase II awards, 12 in all worth $159 million, on the same day.

Models trained on DOE supercomputers are intended to plan pulses in the SPARC control room and support rapid decisions on experimental data as it comes off the machine, compressing hours of heavy simulation into the interval between plasma pulses, according to Tom Looby, the CFS principal investigator leading the project. Brandon Sorbom, CFS Chief Science Officer and co-founder, described the aim as converting decades of research across the US fusion community into tools applied to SPARC operations.

PPPL codes and the digital twin test bed

CFS says the project will be the biggest collaboration to date on applying AI in fusion. It brings together four national laboratories (Lawrence Livermore, Oak Ridge, Princeton Plasma Physics Laboratory and SLAC), two universities (Carnegie Mellon University and the Massachusetts Institute of Technology) and four companies (Google DeepMind, NVIDIA, Synopsys and General Atomics).

Princeton Plasma Physics Laboratory (PPPL) will contribute foundational software engineering and AI models, core physics models, historical tokamak data and high-performance compute, and Michael Churchill, head of its AI for Science Group, leads that work. PPPL’s M3D-C1, TRANSP and XGC simulation codes will help establish full pulse predictions, including through AI surrogates trained on large simulation datasets, which are intended to enable predictions in under 20 minutes, fast enough to use between pulses in a tokamak that confines plasma with magnetic fields.

Digital twin workflows for the National Spherical Torus Experiment-Upgrade (NSTX-U), PPPL’s primary fusion experiment, will serve as a test bed to verify and validate approaches before they are used on SPARC. Some NSTX-U and earlier PPPL data will also be made available to the wider fusion community.

A SPARC digital twin for ARC development

Unlike a conventional simulation, the digital twin will connect to the physical machine and be updated continually, so predictions of device operation improve over time. The integrated workflow between twin and device is intended to optimise and accelerate development of ARC, the commercial machine that follows SPARC, with the Fall Line Fusion Power Station in Chesterfield County, Virginia, as the first ARC plant. DOE describes the Phase II projects as multiyear efforts, and PPPL places validation on NSTX-U ahead of any use on SPARC.

Stay ahead in the fusion revolution explore more breakthroughs from leading innovators in clean energy technology.