Plasma Confinement State Classification in Fusion Power Plants: Profile Reflectometer and Ensemble Diagnostics

This paper presents machine learning classifiers for fusion power plant plasma confinement state using a Profile Reflectometer diagnostic and an ensemble model combining it with Electron Cyclotron Emission data, achieving 97% and 99% test accuracy respectively to address the challenge of limited diagnostic availability in reactor environments.

The Gist

Imagine you are trying to drive a car, but the dashboard is broken. You can't see the speedometer, the fuel gauge, or the engine temperature. All you have is a single, flickering light on the dashboard that tells you if the engine is "running smoothly" or "sputtering." In the world of fusion energy (the technology that aims to replicate the power of the sun), scientists face a similar problem. They need to know the exact state of the super-hot plasma inside a reactor to keep it stable, but future power plants will have very limited space for sensors. They won't be able to install the dozens of complex instruments used in current research labs.

This paper is about teaching a computer to be a "super-driver" that can figure out the engine's condition using only a few, very tough sensors that can survive inside a nuclear reactor.

Here is the story of how they did it, broken down into simple parts:

1. The Goal: Spotting the "High-Performance" Mode

In fusion reactors, there are two main ways the plasma behaves:

• L-Mode (Low): Like a car idling in traffic. It's stable but inefficient.
• H-Mode (High): Like a car speeding down the highway. It's much more efficient and is the goal for future power plants.

The "H-Mode" has a special feature called a pedestal. Think of this like a steep cliff at the edge of the plasma. The temperature and density shoot up right at the edge, creating a barrier that keeps the heat inside. If the computer can spot this "cliff," it knows the reactor is in the good, high-performance mode.

2. The Sensors: Two Different Eyes

The researchers tested two different types of "eyes" (diagnostics) that could survive in a harsh reactor environment:

• The ECE (The Temperature Eye): This sensor looks at the heat (temperature) coming from the plasma. They had already built a smart computer program using this sensor that was pretty good at spotting the H-Mode.
• The PR (The Density Radar): This is the new star of the show. It works like a short-range radar. It shoots radio waves into the plasma and measures how long they take to bounce back. This tells the computer how dense the plasma is at different depths.
  • The Catch: Sometimes, the plasma is so dense that the radar waves can't penetrate all the way to the center. They get stuck at the edge. It's like trying to see through a thick fog; you can see the trees right in front of you, but the mountain in the back is hidden.

3. The Challenge: Dealing with "Foggy" Data

Because the radar (PR) sometimes can't see the center of the plasma, the data is incomplete. The researchers had to teach their computer how to handle this.

• The Solution: Instead of guessing what's in the foggy center, they focused on the edge where the data is clear. They used a mathematical trick (called a "spline") to smooth out the jagged radar lines and create a clean curve. Then, they picked 10 specific points along that curve—mostly focusing on the edge where the "cliff" (pedestal) lives—to feed into the computer.

4. The Results: The Solo vs. The Team

The researchers built three computer models to act as the "driver":

1. The Solo Radar Driver (PR Model): Using only the new radar data, this model was incredibly accurate. It correctly identified the H-Mode 97% of the time. It proved that even with "foggy" data, you can still drive the car if you know where to look.
2. The Solo Heat Driver (ECE Model): This was the previous model using the heat sensor. It was also very good.
3. The Dream Team (Ensemble Model): This is the big innovation. The researchers combined the Radar Driver and the Heat Driver into one "Ensemble" team.
   • How it works: Imagine two navigators in a car. One looks at the heat, the other at the density. If one navigator is confused (because the data is weird or "anomalous"), the other can step in and say, "I'm still clear, trust me." They weigh their answers based on how confident they feel.
   • The Result: This team was nearly perfect, achieving 99% accuracy.

5. Why This Matters for the Future

The paper tested these models not just on random data, but on data that looked like "future experiments" (data the models hadn't seen before).

• Even when the data was tricky or different from the training, the "Dream Team" (Ensemble) held up better than the solo drivers.
• They found that sometimes one sensor sees something weird that the other doesn't. By having both, the system covers each other's "blind spots."

The Bottom Line

This paper shows that we don't need a thousand sensors to run a future fusion power plant. We just need a few tough, reliable ones (like the radar and the heat sensor) and a smart computer that knows how to combine their voices. By teaching the computer to listen to both the "temperature voice" and the "density voice," we can reliably tell if the reactor is running in its most efficient mode, even if the sensors can't see the whole picture perfectly.

In short: They built a smart system that uses two different types of "radar" to tell a fusion reactor when it's in "high gear," proving that even with limited tools, we can keep the future of clean energy running smoothly.

Technical Summary

Technical Summary: Plasma Confinement State Classification Using Profile Reflectometer and Ensemble Diagnostics

Problem Statement
As Fusion Pilot Plants (FPPs) approach feasibility, a critical engineering challenge remains: developing control systems capable of extracting necessary plasma state information from a severely limited subset of diagnostics. Reactor environments impose strict constraints on survivability, maintainability, and port space, rendering many research-focused diagnostics unavailable. Specifically, existing methods for identifying the plasma confinement mode (distinguishing between Low-confinement L-mode and High-confinement H-mode) rely on broad diagnostic suites not viable for future reactors. While the authors previously developed a machine learning (ML) classifier using the Electron Cyclotron Emission (ECE) diagnostic, there is a need to validate and expand this capability using other FPP-viable diagnostics, specifically the Profile Reflectometer (PR), and to determine if combining these modalities improves robustness.

Methodology
The study utilizes data from the DIII-D tokamak, focusing on the Profile Reflectometer (PR) diagnostic, which measures electron density profiles ($n_e$) using Frequency-Modulated Continuous Wave (FM-CW) reflectometry.

1. Data Preparation:

   • The dataset consists of 260 shots (reduced from an initial 300 due to PR data availability and quality constraints) collected across 2024 and 2025.
   • The data is imbalanced, with 66% of time-samples labeled as H-mode and 34% as L-mode.
   • A specific challenge addressed is the "partial coverage" of PR data; due to frequency limits, the diagnostic often fails to penetrate the plasma core ($\rho \approx 0$), capturing only the edge or pedestal region.

2. Feature Extraction (PR Model):

   • To handle noise and variable coverage, the authors fit the raw density profiles using 3rd-order polynomial splines with 10 knots.
   • Instead of extrapolating missing core data (which was deemed less informative for H-mode classification), missing values are padded with the deepest known density value.
   • Ten specific points along the normalized radius ($\rho^*$) are selected as inputs: $[0, 0.2, 0.4, 0.6, 0.8, 0.85, 0.9, 0.95, 1.0, 1.1]$. This distribution places higher density of points near the plasma edge ($\rho \approx 1$), where the H-mode pedestal is located.
   • A Gradient Boosted Classifier (GBC) from sklearn is trained on these spline-fitted values.

3. Ensemble Model:

   • The PR model is combined with the previously developed ECE-based classifier into an ensemble model.
   • The ensemble employs a weighted average approach where the contribution of each model is modulated by a confidence score.
   • Confidence is derived using k-means clustering on the feature space. Data points further from cluster centers (indicating potential anomalies or out-of-distribution samples) are down-weighted.
   • The final prediction is calculated as: $EM = \frac{C_{PR} P_{PR} + C_{ECE} P_{ECE}}{C_{PR} + C_{ECE}}$, where $C$ represents the confidence weight and $P$ the probability output.

Key Results

• PR Classifier Performance: The standalone Profile Reflectometer model achieved a test accuracy of 97% (with a standard deviation of 1.0% across 100 reshuffles). Shapley value analysis confirmed that the model's predictions are driven primarily by the edge region inputs, aligning with the physical intuition that the pedestal defines the H-mode.
• Ensemble Performance: The combined ECE-PR ensemble model achieved a test accuracy of 99.2%.
• Robustness to Future Data: A sliding-window test, designed to simulate temporal shifts in data (training on earlier shots, testing on later shots), showed a performance drop compared to randomized splits but maintained high accuracy. The Ensemble model achieved 96.0% accuracy in this rigorous test, outperforming the individual ECE (91.1%) and PR (93.1%) models.
• Complementarity: Analysis revealed that the two diagnostics often have different "blind spots." When one diagnostic's data point was identified as anomalous (far from the training cluster center), the other diagnostic often remained within the normal range (e.g., when PR was anomalous, 31% of ECE data was normal). This confirms the value of the ensemble approach in mitigating individual diagnostic limitations.

Significance and Claims
The paper claims to establish a benchmark for confinement mode classification using the Profile Reflectometer, a diagnostic considered viable for future fusion reactors due to its compactness, low cost, and ability to operate outside the harsh reactor environment.

The primary significance lies in demonstrating that:

1. High-accuracy plasma state identification is possible using a restricted set of FPP-relevant diagnostics (PR and ECE).
2. An ensemble approach leveraging the complementary nature of these diagnostics can achieve near-perfect accuracy (99%) and greater robustness against data anomalies than single-modality models.
3. The methodology provides a pathway for FPP control systems to operate effectively despite the severe constraints on diagnostic port space and survivability.

The authors conclude that while periodic recalibration may be necessary to assimilate new data as reactor operations evolve, the developed models provide a robust foundation for plasma control in future fusion power plants.