Putting the Sun in a Bottle with AI | AI for Science
At the AI for Science event, Lorenzo Z of the UK Atomic Energy Authority presents work on using AI to accelerate the path to fusion energy, focusing on the STEP spherical tokamak power plant planned to deliver electricity to the grid in the 2040s and backed by £2.5 billion funding to 2030. He explains tokamak fusion, progress toward breakeven, and the need for high fidelity simulation and digital twins that combine messy experimental diagnostics with simulations. He describes three pillars: real time forecasting with multimodal foundation models and FAIR fusion datasets (FAIRMAST); plasma control using simulators such as FreeGSNKE inspired by reinforcement learning advances; and surrogate models that speed up turbulence, edge, and MHD simulations from weeks to hours or milliseconds, enabling design trade offs, uncertainty quantification, active learning to cut data costs, and large compression for simulation storage. Lorenzo Zanisi is a Lead Data Scientist at the UK Atomic Energy Authority (UKAEA), working on one of the most consequential applications of AI in science today: fusion energy. With a PhD in Astrophysics in which he applied AI to galaxy formation simulations, he now leads a team developing surrogate models for fusion plasmas, with pioneering applications to both experiments and reactor design, already delivering speedups of 100x over conventional approaches. 00:00 Welcome and Introduction 01:07 Why Fusion Matters 01:58 How Tokamaks Work 04:44 Progress and Q Factor 06:11 UKAEA and STEP Plan 07:49 Digital Twins and AI Pillars 10:16 Diagnostics and Multimodal Data 12:28 Foundation Model for Plasmas 16:02 FAIR Fusion Dataset FAIRMAST 18:05 AI for Plasma Control 20:54 Surrogate Modeling Workflow 22:45 Turbulence Surrogates 26:27 Trajectory Optimization Faster 28:19 High Fidelity 5D Turbulence 29:50 Cutting Costs and Compression 31:48 Plasma Edge and Wall Heat 34:17 Edge Surrogates for Design 37:28 MHD Instabilities and PDE Models 40:35 What’s Next for AI Fusion 42:45 Thanks and Q&A