DynaBO: Dynamic Model Bayesian Optimization for Tokamak Control

Rohit Sonker, Alexandre Capone, Andrew Rothstein, Hiro Josep Farre Kaga, Egemen Kolemen, Jeff Schneider

Keywords: Nuclear Fusion, Plasma Instabilities, Bayesian Optimization, Applied Machine Learning

TL;DR: Multi Scale Bayesian Optimization which avoids instabilities in nuclear fusion with a success rate of 75% and three fold improvement over past experiments

Abstract: Despite recent advances, state-of-the-art machine learning algorithms struggle considerably with control problems where data is scarce relative to model complexity. This problem is further exacerbated if the system changes over time, making past measurements less useful. While tools from reinforcement learning, supervised learning, and Bayesian optimization alleviate some of these issues, they do not address all of them at once. Considering these drawbacks, we present a multi-scale Bayesian optimization for fast and data-efficient decision-making. Our pipeline combines a high-frequency data-driven dynamics model with a low-frequency Gaussian process, resulting in a high-level model with a prior that is specifically tailored to the dynamics model setting. By updating the Gaussian process during Bayesian optimization, our method adapts rapidly to new data points, allowing us to process current high-quality data quickly, which is more representative of the system than past data. We apply our method to avoid tearing instabilities in a tokamak plasma, a control problem where modeling is difficult, and hardware changes potentially between experiments. Our approach is validated through offline testing on historical data and live experiments on the DIII-D tokamak. On the historical data, we show that our method outperforms a naive decision-making algorithm based exclusively on a recurrent neural network and past data. The live experiment corresponds to a high-performance plasma scenario with a high likelihood of instabilities. Despite this base configuration, we achieved a 50\% success rate in the live experiment, representing an improvement of over 117\% compared to historical data.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 12548