Abstract
Controlling nuclear fusion plasmas is notoriously difficult due to complex, shifting dynamics and the prohibitive cost of running experiments. This paper introduces DynaBO, a multi-timescale Bayesian optimization framework that combines a fast neural network dynamics model with a slow-updating Gaussian process to stabilize plasmas. Tested live on the DIII-D tokamak, the method successfully suppressed tearing instabilities in 50% of highly unstable scenarios, marking a 117% improvement over historical baselines.
Introduction
Nuclear fusion via tokamaks promises clean, limitless energy, but the technology is plagued by plasma disruptions. A major culprit is the "tearing mode instability"—a phenomenon where magnetic islands form in the plasma, leading to energy loss and potential damage to the reactor walls. Standard machine learning approaches struggle here; model-free reinforcement learning requires thousands of trial-and-error rollouts, which is financially and physically impossible on a real reactor. Furthermore, tokamak dynamics are a moving target: the reactor's behavior changes slowly over time due to hardware wear, wall conditioning, and impurities. The authors needed a highly data-efficient way to find optimal feedforward control profiles for Electron Cyclotron Heating (ECH) using only a handful of live experimental trials.
Methodology
To bridge the gap between complex physics and limited live data, the authors propose DynaBO, a multi-scale contextual Bayesian Optimization approach. The framework consists of a high-frequency Recurrent Probabilistic Neural Network (RPNN) trained on historical data to predict plasma state transitions, and a low-frequency Gaussian Process (GP) adapter. The GP predicts the time-to-tearing mode based on the target plasma pressure and the ECH profile, using the RPNN's predictions as the prior mean.
Crucially, DynaBO utilizes a hardware-agnostic action space by optimizing the shape of the heating profile (parameterized as a Gaussian curve) rather than raw hardware knobs. This allows the system to remain robust even if specific hardware components become unavailable. Between-shot updates using Upper Confidence Bound (UCB) acquisition allow the model to adapt to the reactor's current state on the fly.
Experimental Results
Offline simulations were conducted using historical data from the DIII-D tokamak (2010–2023). DynaBO achieved significantly lower cumulative regret compared to baselines, including RPNN-only models and vanilla GPs with zero or mean priors. This demonstrates superior exploration capabilities and robustness across various kernel choices.
In live experimental shots on the DIII-D tokamak, DynaBO was tested on a highly unstable "high $q_{min}$" plasma configuration. Historically, this configuration suffers a 77% tearing mode instability rate (a 23% success rate). DynaBO successfully avoided tearing modes in 4 out of 8 shots, representing a 50% success rate and a >117% relative improvement. Furthermore, DynaBO increased the average stable plasma duration to 3339 ms, a 914 ms improvement over the historical average of 2424 ms.
Conclusion and Future Work
DynaBO demonstrates that multi-scale modeling is highly effective for drifting environments. By using a neural network as a prior for a GP, the system can quickly adapt to the current state of the reactor using just a handful of live shots. While the current method focuses on feedforward control, future work aims to extend this to real-time feedback control and incorporate physics-informed neural networks to further improve interpretability and data efficiency.