Choose timezone
Your profile timezone:
CD Technical Meeting (ML7): AI for Fusion: Old Challenges, New Tools
→
Europe/London
Description
Machine Learning, Uncertainty Quantification and Data Science
To be presented at the AI for Science Symposium, The Royal Swedish Academy of Sciences, Stockholm, September 4th, 2025
Nuclear fusion represents humanity's most promising pathway to clean, abundant energy, yet achieving controlled fusion has remained elusive due to the extreme physics involved—requiring temperatures of 150 million degrees Celsius and precise confinement of volatile plasma within complex magnetic field geometries. This talk demonstrates how artificial intelligence is fundamentally transforming fusion research by addressing these decades-old challenges through four key applications: deploying neural operators and surrogate models that accelerate computationally intensive magnetohydrodynamic simulations by up to six orders of magnitude; creating digital twins for real-time prediction of plasma instabilities and behavioral evolution; implementing reinforcement learning algorithms that autonomously control thousands of plasma parameters while maintaining stability, as demonstrated in recent breakthroughs by Google DeepMind; and employing multi-objective Bayesian optimization to efficiently explore high-dimensional design spaces for next-generation tokamak reactors like the UK's STEP program. Unlike traditional engineering approaches that rely on iterative physical prototyping, fusion's prohibitive experimental costs necessitate an AI-driven, simulation-first methodology that bridges the critical gap between computational models and experimental reality through robust uncertainty quantification and validated digital representations of plasma behavior.
