Enabling Integrated AI Control on DIII-D: A Control System Design with State-of-the-art Experiments

TL;DR

The paper presents PACMAN, a modular, real-time framework for Prediction And Control using MAchiNe learning, implemented on the DIII-D tokamak to enable end-to-end ML-driven control experiments. By separating diagnostic inputs, model computations, controllers, and actuation post-processing across a three-level design, PACMAN achieves scalable, fault-tolerant operation that can integrate diverse ML models (RL, survival models, RC networks) and standard controllers (PID, FSM, MPC). Five experimental deployments demonstrate the framework's versatility, including RL-based beta_N/ITB control, TM suppression with ECCD, ELM prediction, AE suppression, and latent-space MPC for profile control, with cycle times from sub-ms to tens of ms. The work provides design principles and an appendix with practical planning guidance to extend PACMAN to future diagnostics, actuators, and hardware accelerators, signaling a path toward broader adoption of integrated ML control in fusion devices.

Abstract

We present the design and application of a general algorithm for Prediction And Control using MAchiNe learning (PACMAN) in DIII-D. Machine learing (ML)-based predictors and controllers have shown great promise in achieving regimes in which traditional controllers fail, such as tearing mode free scenarios, ELM-free scenarios and stable advanced tokamak conditions. The architecture presented here was deployed on DIII-D to facilitate the end-to-end implementation of advanced control experiments, from diagnostic processing to final actuation commands. This paper describes the detailed design of the algorithm and explains the motivation behind each design point. We also describe several successful ML control experiments in DIII-D using this algorithm, including a reinforcement learning controller targeting advanced non-inductive plasmas, a wide-pedestal quiescent H-mode ELM predictor, an Alfvén Eigenmode controller, a Model Predictive Control plasma profile controller and a state-machine Tearing Mode predictor-controller. There is also discussion on guiding principles for real-time machine learning controller design and implementation.

Figures (5)

• Figure 1: DIII-D diagnostics, actuators and PACMAN. PACMAN receives diagnostic data and finds the optimal actuators to accomplish the user defined control targets. Location of diagnostics and actuators are approximate in this sketch.
• Figure 2: General design of the PACMAN algorithm. The algorithm is a end-to-end controller, starting from diagnostic inputs, followed by ML, standard control, or physics-based models, then saving the ML outputs and sending the data to a controller such as finite state machines and PIDs, and ending with commands to actuators such as neutral beam injector (NBI) power, electron cyclotron heating (ECH) power, and gas requests.
• Figure 3: 3 level design structure of the PCS diagnostics and algorithms. The flow of diagnostic information is always up levels and to the right, where raw diagnostic data is processed, then may be fed to automated fitting routines such as real-time EFITferron_real_1998, before being provided as inputs to various controllers. On the controller side, commands travel down the levels from the highest controllers, which may decide on target profile objectives, or plasma stability targets, and gets passed down to lower level controllers that will achieve those high level goals before the lowest level controllers calculate exact commands to be send to hardware. Of note, as seen in the figure there are typically more Level 1 diagnostics and controllers than there are Level 2 or 3. PACMAN is at the top of the pyramid as a Level 3 controller.
• Figure 4: Overview of the 5 experimental applications that are described. Each block is depicted mirroring the layout from Figure \ref{['fig:general_design']}, where each horizontal dashed line separates each experiment. Note the many shared inputs and outputs that utilize the framework.
• Figure 5: DIII-D shot 204975 where the RL ITB controller was used to control $\beta_N$ and ITBs. On the left are three of the real-time inputs: $\beta_N$, core electron temperature, and the $q$ profile at $\psi_N=0.95$. These along with the remaining inputs are fed into the model where the mdoel outputs information on which RL model was selected as well as ITB strength and width. This information is fed into the controller block that finalizes the ECH, on-axis NBI power, and off-axis NBI power that get sent to the actuators.