Control of pedestal-top electron density using RMP and gas puff at KSTAR

This paper reports the successful development and experimental validation of a real-time, multi-actuator controller at KSTAR that utilizes resonant magnetic perturbations and gas puffing to precisely control pedestal-top electron density with high accuracy and dynamic responsiveness.

The Gist

Imagine you are trying to keep a pot of soup at the perfect temperature while cooking a very delicate dish. If the soup gets too hot, it boils over and ruins the meal; if it's too cold, it never cooks. In the world of nuclear fusion, the "soup" is a super-hot ball of plasma (electrically charged gas) floating inside a magnetic donut-shaped machine called a tokamak.

The scientists at KSTAR (a fusion experiment in South Korea) faced a similar problem. They needed to keep the density of electrons in a specific "crust" of the plasma (called the pedestal) at just the right level. If this density is wrong, the plasma becomes unstable, or worse, it might damage the walls of the machine.

Here is how they solved it, explained through simple analogies:

1. The Problem: A Moving Target

The goal was to keep the electron density at a specific spot steady, even as the plasma changed. But it wasn't just about staying still; they wanted to be able to change the density on purpose, like turning a dial up or down, to test different cooking recipes (experimental scenarios).

Previously, they had to guess and check manually, which is slow and risky. They needed a robot chef that could taste the soup and adjust the heat instantly.

2. The "Eyes": Seeing the Invisible

To control the soup, you need to see it. But you can't stick a thermometer into a million-degree plasma; it would melt instantly.

• The Solution: They used a special laser system (a two-colored interferometer) that shoots beams through the plasma. By measuring how much the light bends, they can calculate how "thick" the plasma is.
• The Speed Bump: Doing the math to turn those light measurements into a 3D map of the plasma used to take seconds. For a robot chef, that's an eternity. By the time the computer finished the math, the soup would have already burned.
• The AI Fix: The team trained a neural network (a type of simple AI) to do the math. Think of it as teaching a student to recognize patterns so fast they can guess the answer in a blink. Instead of taking seconds, the AI now calculates the plasma density in 120 microseconds (that's 0.00012 seconds). It's fast enough to be used in real-time control.

3. The "Hands": Two Different Tools

To adjust the density, the scientists had two tools, but they worked in opposite ways:

• Tool A (The Vacuum Cleaner): They used Resonant Magnetic Perturbations (RMP). Imagine wiggling the magnetic field slightly. This acts like a vacuum cleaner, sucking some particles out of the plasma to lower the density.
• Tool B (The Water Hose): They used a gas puff. This injects fresh gas (Deuterium) into the plasma to raise the density.

4. The "Brain": The Smart Controller

The real magic was the controller (the brain of the operation).

• The Challenge: If you only use the vacuum cleaner, you can only lower the density. If you only use the hose, you can only raise it. You can't do both at once, or you'd just be fighting yourself.
• The Solution: They built a "Proportional-Integration" (PI) controller. Think of this as a smart thermostat.
  • If the density is too high, the brain tells the vacuum cleaner (RMP) to turn on and the hose to turn off.
  • If the density is too low, the brain tells the hose to spray gas and the vacuum to turn off.
  • Crucially, the brain ensures these two tools never fight each other. It's like a traffic light that ensures cars only go one way at a time, preventing a crash.

5. The Result: A Master Chef

The team tested this system in the 2024-2025 experiments.

• Accuracy: The system kept the density within 1.5% of the target. That's like hitting a bullseye on a dartboard from across the room.
• Flexibility: They could ask the system to lower the density, then raise it, then lower it again, all while the machine was running. The system followed these "dynamic targets" perfectly.
• Speed: Because the AI was so fast, the controller could react instantly to changes, keeping the plasma stable even when things got turbulent.

Why Does This Matter?

In the future, we want fusion power plants to run 24/7. To do that, the plasma must be perfectly stable and "detached" (not touching the walls). This new controller is like a autopilot for the fusion reactor. It allows scientists to quickly test different "recipes" (density levels) to find the perfect conditions for clean, limitless energy without having to stop the machine and start over.

In short: They taught a computer to "see" the plasma instantly using AI, and then gave it a smart brain to juggle two opposite tools (a vacuum and a hose) to keep the plasma at the perfect density, all in the blink of an eye.

Technical Summary

Here is a detailed technical summary of the paper "Control of pedestal-top electron density using RMP and gas puff at KSTAR."

1. Problem Statement

Future tokamak fusion reactors require High-Confinement Mode (H-mode) operation to achieve necessary plasma performance. However, H-mode is prone to Edge Localized Modes (ELMs), which cause damaging heat and particle bursts on the divertor, and requires detachment to manage heat loads. Both ELM suppression and detachment are highly sensitive to the pedestal-top electron density ($n_e$).

• The Challenge: To achieve an ELM-free or detached state, the pedestal-top $n_e$ must be maintained within a specific, narrow operational window. Furthermore, to explore these regimes efficiently, the density must be controllable in real-time to follow dynamic targets (increasing or decreasing) within a single plasma discharge (shot).
• Limitations of Previous Methods: Offline profile reconstruction is too slow for real-time control. Existing control strategies often rely on single actuators (either Resonant Magnetic Perturbations or gas puffing), which cannot independently increase and decrease density to follow arbitrary dynamic targets.

2. Methodology

The authors developed a closed-loop control system implemented in the KSTAR Plasma Control System (PCS) to regulate the electron density at $\psi_N = 0.89$ (the pedestal top). The methodology consists of three main components:

A. Real-Time Electron Density Profile Reconstruction

• Data Sources: The system combines equilibrium reconstruction data from EFIT (specifically the normalized poloidal flux grid, $\psi_N$) and line-averaged density measurements from the Two-Colored Interferometer (TCI) (5 channels).
• Parametric Model: A fitting function models the $n_e$ profile using three regions:
  1. Pedestal: Hyperbolic tangent function.
  2. Core: Polynomial function.
  3. Scrape-Off Layer (SOL): Linear function.
     The model uses 6 parameters, but 2 are fixed to reduce complexity given the limited TCI channels.
• Neural Network Acceleration: The traditional iterative fitting process takes milliseconds, which is too slow for the 500 $\mu$s PCS cycle time. The authors trained a Multi-Layer Perceptron (MLP) to replace the iterative fitting.
  • Inputs: 10 variables (flux mapping coefficients, geometric radii, and 5 TCI density measurements).
  • Outputs: 4 fitting coefficients ($a_1$ to $a_4$).
  • Performance: The MLP runs in ~120 $\mu$s, enabling real-time reconstruction with a median absolute percentage error of 1.85% on test data.

B. System Identification

To design the controller, the plasma response to actuators was modeled as a first-order linear time-invariant (LTI) system:
$$T \frac{dy}{dt} + y = K u(t)$$

• Actuators:
  1. RMP (Resonant Magnetic Perturbation): Applied via In-Vessel Control Coils (IVCC). Effect: Reduces density (pump-out).
  2. Main Gas Puff: Injects $D_2$ gas. Effect: Increases density.
• Identification: The gain ($K$) and time constant ($T$) were estimated by minimizing the error between the experimental response and the model convolution. Notably, the system identification for the main gas puff was approximated using data from a bottom valve (PVB) due to experimental constraints.

C. Controller Design

• Algorithm: A Proportional-Integral (PI) controller was selected.
• Gain Calculation: Gains were determined using pole placement. The closed-loop poles were placed to match the plant's fast poles to ensure robustness.
• Mutually Exclusive Logic: Since RMP decreases density and gas puff increases it, the controller was designed to use only one actuator at a time. This was achieved by exploiting the sign difference in the control gains ($K_p$ and $K_I$) for the two actuators. If the error requires a density drop, the RMP command is positive and gas is zero; if a rise is needed, gas is positive and RMP is zero.

3. Key Contributions

1. Real-Time Profile Reconstruction: Successfully deployed an MLP to accelerate the $n_e$ profile reconstruction from the order of milliseconds to 120 $\mu$s, making it compatible with the KSTAR PCS.
2. Dual-Actuator Exclusive Control: Developed a control logic that seamlessly switches between RMP and Main Gas Puff to both increase and decrease pedestal density, allowing for dynamic target tracking (ramps up and down) within a single shot.
3. Generalizable Gains: Demonstrated that controller gains derived from system identification on specific configurations (e.g., $n=1$ RMP, PVB gas) are generalizable to different plasma currents and actuator configurations (e.g., $n=2$ RMP, Main Gas Puff).

4. Experimental Results

The controller was tested during the 2024-2025 KSTAR campaign:

• Single Actuator (RMP only):

  • Successfully tracked a dynamic target (linear decrease then increase) using only RMP.
  • Accuracy: Median absolute percentage error of 1.72%; average error of 2.72%.
  • Observed density reduction correlated with RMP current application and a drop in normalized beta ($\beta_N$).

• Dual Actuator (RMP + Main Gas Puff):

  • Successfully tracked a more aggressive dynamic target, utilizing RMP to lower density and Main Gas Puff to raise it.
  • Accuracy: Median absolute percentage error of 1.64%; average error of 2.20%.
  • The system maintained mutual exclusivity of actuators as designed, preventing conflicting commands.
  • The controller proved robust across different plasma currents (0.5 MA vs. 0.7 MA) and RMP mode numbers ($n=1$ vs. $n=2$).

5. Significance

• Operational Flexibility: This system enables "scenario exploration" within a single discharge. Researchers can now dynamically scan the pedestal density to find the optimal window for ELM suppression or detachment without resetting the plasma.
• Efficiency: Reduces the need for multiple experimental shots to map out density dependencies, saving valuable machine time.
• Foundation for Advanced Control: This work lays the groundwork for integrated control systems that combine density regulation with ELM suppression and detachment controllers, which is critical for the operation of future fusion power plants like ITER and DEMO.
• Machine Learning Integration: Demonstrates the practical viability of using lightweight neural networks for real-time physics reconstruction in high-performance tokamak control systems.