New AI method accelerates plasma heat defense in reactors

Researchers from Commonwealth Fusion Systems, the DOE’s Princeton Plasma Physics Laboratory, and Oak Ridge National Laboratory have developed a new AI method called HEAT-ML to accelerate the protection of fusion reactors from extreme plasma heat. HEAT-ML enhances the existing Heat flux Engineering Analysis Toolkit (HEAT) by using a deep neural network trained on about 1,000 SPARC tokamak simulations to rapidly generate 3D “shadow masks.” These masks identify regions of the reactor’s inner walls shielded from direct plasma contact, which is critical to preventing damage from plasma temperatures exceeding those at the Sun’s core. Traditional HEAT simulations can take up to 30 minutes per run, whereas HEAT-ML produces results in milliseconds, dramatically speeding up the design and operational decision-making processes for fusion systems. The AI was initially tested on 15 tiles near the bottom of SPARC’s exhaust system, the area expected to experience the highest heat loads. By quickly and accurately locating magnetic shadows, HEAT-ML supports