Building a validated Digital Twin for Fusion

In this talk, we will describe an AI-enabled digital twin of the DIII-D National Fusion Facility that transforms fusion energy research by reducing plasma simulation times from weeks to seconds. This effort is a collaboration between NVIDIA, General Atomics, and research partners including San Diego Supercomputer Center, Argonne Leadership Computing Facility, and NERSC. The project addresses fusion's critical challenge of controlling plasma at hundreds of millions of degrees while predicting its behavior rapidly enough to maintain reactor stability. Built within NVIDIA Omniverse and powered by GPU-accelerated AI surrogate models—including EFIT for plasma equilibrium, CAKE for plasma boundary prediction, and ION ORB for ion heat density—the digital twin integrates real-time sensor data from the physical reactor with physics-based simulations and engineering models. Trained using the Polaris and Perlmutter supercomputers on decades of experimental data, this interactive virtual environment enables 700 scientists from over 100 organizations to conduct risk-free "what-if" scenario testing and rapidly iterate reactor designs without endangering physical infrastructure. This "fusion accelerator" represents a paradigm shift from purely physics-driven experimentation to AI-augmented computational discovery, significantly accelerating the path toward commercial fusion energy.