Nvidia and General Atomics deliver digital twin for fusion energy research (blogs.nvidia.com)

NVIDIA, General Atomics and national lab partners announced an AI-enabled, high-fidelity digital twin of the DIII-D fusion reactor that runs interactively in near real time. Using Polaris (ALCF) and Perlmutter (NERSC) to train three large AI surrogate models at scale—EFIT (plasma equilibrium), CAKE (plasma boundary) and ION ORB (heat density of escaping ions)—the project combines NVIDIA Omniverse, CUDA-X libraries, RTX PRO servers and DGX Spark with data and compute from the San Diego Supercomputer Center, ALCF and NERSC. The digital twin fuses sensor streams, physics-based simulations, engineering models and AI surrogates, synchronized with the physical reactor so 700 scientists across 100 organizations can run “what-if” experiments without touching the machine. The technical payoff is dramatic: simulations that once took weeks can now deliver actionable predictions in seconds, enabling real-time control strategies to stabilize plasma—an essential step for preventing damage and accelerating experiments. Trained on decades of experimental data, the surrogates trade some physics-model complexity for massive speed gains, making interactive exploration, rapid optimization and safer procedure development practical. Beyond DIII-D, this work reframes fusion progress as a cross-disciplinary computing challenge, showing how GPU-accelerated AI and digital twins can materially shorten the path toward commercially viable fusion energy.