Abstract
Managing heat exhaust in fusion reactors is a critical challenge for the longevity of reactor components. This study introduces a "white-box" AI control system designed to regulate divertor detachment in the DIII-D tokamak. By utilizing a linear regression model trained on high-fidelity machine learning labels, the researchers achieved real-time plasma control with a mean absolute error of only 2%. This interpretable framework provides a transparent alternative to opaque deep learning models, meeting the stringent safety and auditability requirements likely to be imposed on future commercial fusion power plants.
Introduction
In a tokamak fusion reactor, the divertor serves as the primary exhaust system, absorbing intense heat from the plasma. To prevent structural damage, it is essential to achieve a state known as detachment. In this state, a cushion of cool gas forms near the divertor wall, dissipating heat before it reaches the surface. However, controlling detachment is difficult due to the highly non-linear behavior of plasma on millisecond timescales.
While deep learning models like Convolutional Neural Networks (CNNs) have shown high performance in predicting plasma states, their "black-box" nature poses a significant hurdle for regulatory approval in commercial energy production. This research addresses the need for "interpretable" AI that combines high-performance prediction with the transparency required for safety audits.
Methodology and Interpretable ML
The core of the proposed system is a real-time feedback control loop that relies on the TangTV diagnostic system. TangTV uses a tangential camera to capture Carbon-III (C-III) light emissions, which serve as a proxy for the plasma's temperature and detachment state.
The Linear Model
Instead of a complex deep neural network, the researchers opted for a linear regression model. This choice ensures that the model's decision-making process is transparent; the model essentially performs a dot product between the camera image and a weight map. This allows engineers to visualize exactly which regions of the plasma (e.g., the emission "leg" or "ring") the AI is prioritizing, ensuring it is not reacting to background noise or artifacts.
The Control Metric ($DZ$)
The system uses a normalized metric, $DZ$, to quantify the detachment state:
- $DZ = 0$: Plasma is in contact with the wall (strike point).
- $DZ = 1$: Plasma is fully detached up to the X-point (the magnetic null point).
To train this linear model, the team used "ground truth" labels derived from 3,500 tomographically inverted images. While tomographic inversion is too computationally expensive for real-time use, it provides the high-fidelity data needed to train a simpler, faster linear model for live operations.
Real-Time Control Architecture
The inferred $DZ$ signal is integrated into a standard Proportional-Integral-Derivative (PID) controller. This controller adjusts the injection of Deuterium ($D_2$) gas to maintain the desired $DZ$ level.
A key innovation in this architecture is the inclusion of a geometric correction factor ($Z_{adj}$). By incorporating real-time magnetic equilibrium data from rt-EFIT, the controller remains robust against radial shifts in the plasma position that might otherwise misinform the image-based AI.
Experimental Results and Validation
The system was validated on the DIII-D tokamak across various plasma scenarios, including both L-mode (low confinement) and H-mode (high confinement). Key results include:
- Precision: The feedback loop maintained the target detachment level with a mean absolute difference of just 2%.
- Physical Consistency: Independent diagnostics confirmed the AI's performance. When the $DZ$ metric reached 1.0, infrared sensors recorded a significant drop in heat flux to the divertor (approximately 0.55 $MW/m^2$), validating that the AI-inferred state corresponded to physical detachment.
Conclusion and Future Directions
This work demonstrates that complex machine learning can be distilled into simple, interpretable linear systems for high-stakes real-time control. While successful, several areas for improvement remain:
- Latency: Current system delays (approx. 250ms total) could be reduced through GPU acceleration.
- Impurity Handling: Future reactors using tungsten walls will require the model to be adapted for different emission spectra (e.g., Neon or Argon).
- ELM Mitigation: Integrating techniques to handle Edge Localized Modes (ELMs)—which can momentarily "blind" the camera—will be crucial for sustained operation in high-performance regimes.
By bridging the gap between advanced AI and classical control theory, this research provides a viable path for regulation-compliant automation in the next generation of fusion power plants.
Summary updated