Speaker: Dr. Michael Churchill, Princeton Plasma Physics Lab
Title: AI to accelerate fusion energy device design, operation, and discovery
Abstract:
Fusion energy devices couple complex physics and engineering across a wide range of scales, making both high-fidelity simulation and experimentation expensive. Artificial intelligence offers a path to combine simulation, experimental data, and reduced predictive models to accelerate device design, improve operation, and enable scientific discovery. Here I will touch on a variety of fusion energy projects that leverage AI.
The first is optimizing stellarator designs to minimize turbulence transport. Instead of the traditional route using physics-based proxies, we develop AI surrogates based on a large database of nonlinear, local gyrokinetic turbulence simulations for high(er)-fidelity direct calculation of turbulent heat flux. We demonstrate the expanded possibilities for optimization targets using the AI surrogate, including prediction across the entire plasma and critical gradient optimization. We also demonstrate using AI agents to drive large scale optimization campaigns with these surrogates, intelligently exploring the design space running hundreds of DESC optimizations for magnetic configurations with optimally reduced turbulent transport, with leading candidates showing a significant >40% increase in core ion temperature.
The second is creating AI-based digital twins for operational excellence in the upcoming NSTX-U spherical tokamak. These digital twins integrate simulations, diagnostic measurements, and AI models to provide the capability to better extrapolate to unseen conditions, such as upgrade machines like NSTX-U. These capabilities are being connected through an action-conditioned AI-based world model that learns plasma-state evolution under changing actuator trajectories, enabling rapid exploration of possible operating scenarios, uncertainty quantification, and between-shot optimization.
Together, these efforts illustrate a broader strategy for AI in fusion: use high-fidelity physics to train and constrain learned models, use experiments to continually correct them, and deploy the resulting model hierarchy to explore design spaces and operational trajectories that would otherwise be computationally inaccessible.
Bio: Dr. Michael Churchill is the Head of AI for Science and Digital Engineering in the Computational Science Department of the Princeton Plasma Physics Laboratory (PPPL), whose research focuses on integrating physics and engineering modeling/simulation for magnetic confinement fusion energy systems. Dr. Churchill earned a Ph.D. degree in Nuclear Science and Engineering in 2014 at the Massachusetts Institute of Technology (MIT), performing experimental research on the Alcator C-Mod tokamak. He received his BA in Electrical and Computer Engineering from Brigham Young University.
----
Seminar Access: In-person seminars are only available to CU ID holders. Non-Columbia affiliates can contact [email protected] if you'd like the seminar Zoom link.