Plasma Confinement State Classification via FPP Relevant Microwave Diagnostics

Basic Information

• Paper ID: 2510.14078
• Title: Plasma Confinement State Classification via FPP Relevant Microwave Diagnostics
• Authors: Randall Clark, Vacslav Glukhov, Georgy Subbotin, Maxim Nurgaliev, Aleksandr Kachkin, Max Austin, Dmitri M. Orlov
• Classification: physics.plasm-ph
• Publication Date: August 2025 (arXiv preprint)
• Paper Link: https://arxiv.org/abs/2510.14078

Abstract

This paper proposes a concise and robust machine learning approach for identifying plasma confinement states in fusion power plants (FPPs), where reliable discrimination between low confinement mode (L-mode) and high confinement mode (H-mode) is critical for safe and efficient operation. Unlike research-oriented devices, FPPs must operate under severely limited diagnostic equipment constraints. To address this challenge, the authors demonstrate that a minimalist model using only electron cyclotron emission (ECE) signals can provide accurate and reliable state classification. ECE provides electron temperature profiles without facing engineering or survivability issues associated with in-vessel probes, making it a prime candidate for FPP-relevant diagnostics. The framework employs ECE as input, uses radial basis functions for feature extraction, and applies gradient boosting classifiers, achieving high accuracy with average test accuracy of 96%. Robustness analysis and feature importance studies confirm the reliability of the method.

Research Background and Motivation

1. Core Problem

The core problem addressed by this research is how to accurately identify plasma confinement states (L-mode vs H-mode) using only limited diagnostic equipment in the harsh environment of fusion power plants (FPPs).

2. Problem Significance

• Safe Operation Requirements: H-mode is the primary operating mode for FPPs because it forms edge transport barriers and pedestals, increasing density and temperature while reducing energy losses
• Engineering Constraints: The FPP environment is extremely harsh, including neutron bombardment, high heat and particle fluxes, strong electromagnetic forces, and γ-radiation, all of which erode sensors
• Maintenance Limitations: Physical access is restricted, making it difficult to repair damaged equipment

3. Limitations of Existing Methods

• Research tokamaks (such as DIII-D) possess over 50 diagnostic systems, but most are infeasible in FPP environments
• Existing H-mode classifiers rely on extensive diagnostic equipment sets that are unavailable in reactor environments
• Traditional methods lack robustness to diagnostic equipment failures

4. Research Motivation

To bridge the technological gap in plasma state classification for future fusion devices under diagnostic constraints, and to develop minimalist yet resilient plasma control architectures for FPPs.

Core Contributions

1. Proposed a minimalist ECE-based classification framework: Demonstrated that high-precision L/H-mode classification can be achieved using only electron cyclotron emission signals
2. Developed physics-driven feature extraction methods: Employed radial basis functions (RBF) to capture physical information in temperature profiles, particularly H-mode pedestal characteristics
3. Achieved high classification performance: Test accuracy of 96% and F1 score of 95%
4. Provided comprehensive robustness analysis: Assessed the impact of ECE channel failures and noise on model performance
5. Constructed high-quality annotated dataset: Based on 300 discharges from DIII-D with 7,668 expert-annotated data points

Methodology Details

Task Definition

Input: 40-channel ECE electron temperature profile data, supplemented by CO2 interferometer density measurements and magnetic diagnostics Output: Binary classification result (L-mode or H-mode) Constraints: Use only diagnostic equipment that can survive in FPP environments

Model Architecture

1. Diagnostic System Design

• ECE System: Provides electron temperature profile Te(R,t), operating in the ~100 GHz microwave range
• CO2 Interferometer: Provides line-integrated electron density, serving as a validity check for ECE
• Magnetic Diagnostics: Toroidal magnetic field BT and plasma boundary position Rout_mid

2. Data Preprocessing Pipeline

Raw ECE data → Density limit check → Pedestal observation check → RBF feature extraction → Classifier

3. Radial Basis Function Feature Extraction

Temperature profiles are interpolated using Gaussian RBF:

f(x) = Σ(q=1 to Nc) ωq * φ(||x - c(q)||)
φ(||x - c(q)||) = exp(-k(||x - c(q)||)²)

Where:

• Nc = 7 center points
• k = 10²
• Center points uniformly distributed from plasma core to boundary

4. Gradient Boosting Classifier

Employs sklearn's HistGradientBoostingClassifier with key parameters:

• Learning rate: 0.1
• Maximum iterations: 100
• Maximum leaf nodes: 31

Technical Innovations

1. Physics-driven feature engineering: RBF weights directly encode physical information in temperature profiles, particularly H-mode pedestal characteristics
2. Minimalist diagnostic set: Demonstrated that ECE alone is sufficient for accurate classification, significantly simplifying FPP diagnostic requirements
3. Engineering constraint consideration: All selected diagnostics have been assessed for survivability in FPP environments
4. Robustness design: Model can handle partial ECE channel failures and noise interference

Experimental Setup

Dataset

• Source: DIII-D tokamak experimental data from 2024-2025
• Scale: 283 discharges, 7,668 data points
• Annotation: Expert manual annotation, 37% H-mode, 63% L-mode
• Time Resolution: Sampled at 100ms intervals
• Quality Control:
  • Excluded data points with density exceeding 4.5×10¹³ cm⁻³
  • Excluded data points where ECE channels cannot observe the pedestal
  • Edge distance parameter R_Δedge = 0.1m

Evaluation Metrics

• Test Accuracy
• Precision
• Recall
• F1 Score

Comparison Methods

• Decision stump classifier using single RBF weights (for feature importance analysis)
• Traditional classifiers including decision trees and support vector machines

Implementation Details

• Train/test split: 80/20 by discharge number (avoiding temporal leakage)
• Cross-validation: 100 shuffled repeats
• Density cutoff: 4.5×10¹³ cm⁻³
• Temperature threshold: 90% of maximum temperature

Experimental Results

Main Results

Feature Importance Analysis

ROC curve analysis reveals that RBF weights closest to the plasma edge (weights 6 and 7) are most discriminative for L/H-mode classification, consistent with H-mode pedestal physics mechanisms.

Model Validation

Testing on 4 independent discharges outside the training set:

• Shot 188859: High-confidence H-mode classification, with confidence decreasing only when density exceeds limits
• Shot 198759: Correctly identified L-H and H-L transitions, but with reduced confidence during strong oscillations
• Shot 199850: Accurately detected multiple L-H/H-L transitions
• Shot 200324: Reconfirmed the impact of density limits on model accuracy

Robustness Analysis

1. ECE Channel Failure Tolerance

• Model maintains high accuracy with 15 or more channels
• Performance degrades sharply below 15 channels
• Edge channels are more important than core channels

2. Noise Robustness

Tested using multiplicative noise model:

T̂ = T * exp(N(-σ²/2, σ²))

Results show the model maintains good performance for σ < 0.2.

Experimental Findings

1. Feasibility of minimal diagnostic set: ECE alone achieves state-of-the-art performance
2. Physical consistency: High importance of edge RBF weights aligns with H-mode pedestal physics mechanisms
3. Engineering practicality: Model demonstrates good tolerance to common diagnostic failures

Related Work

Main Research Directions

1. Fusion Diagnostics: ECE, Thomson scattering, magnetic diagnostics, etc.
2. Plasma Control: Shape control, divertor detachment prediction
3. ML Applications in Fusion: H-mode detection, safety factor reconstruction, disruption prediction

Advantages Over Related Work

• Diagnostic Constraints: First work specifically designed for diagnostic constraints in FPP environments
• Minimalism: Requires only ECE compared to existing methods requiring multiple diagnostics
• Robustness: Provides comprehensive failure tolerance analysis
• Physics-driven: RBF feature extraction based on H-mode physics mechanisms

Conclusions and Discussion

Main Conclusions

1. Technical Feasibility: Demonstrated that 96% H-mode classification accuracy can be achieved using only ECE
2. Engineering Applicability: Selected diagnostics have good survivability in FPP environments
3. Robustness: Model demonstrates good tolerance to diagnostic failures and noise
4. Physical Consistency: Feature importance aligns well with H-mode physics mechanisms

Limitations

1. Single-device Validation: Verified only on DIII-D; portability requires testing on other devices
2. Density Limitations: ECE signal quality degrades at high densities, affecting performance
3. Boundary Reconstruction Dependency: Still requires accurate plasma boundary information
4. Annotation Subjectivity: Expert annotation may contain some subjectivity

Future Directions

1. Multi-diagnostic Fusion: Integrate other FPP-compatible diagnostics such as reflectometry and polarimetry
2. Hybrid Framework: Develop multi-diagnostic fallback mechanisms to enhance overall robustness
3. Real-time Implementation: Optimize algorithms to meet real-time control requirements
4. Cross-device Validation: Verify on other devices such as ITER and SPARC

In-Depth Evaluation

Strengths

1. Strong Engineering Orientation: Specifically designed for practical FPP constraints with high practical value
2. Physics-driven Methodology: RBF feature extraction directly corresponds to H-mode physics mechanisms
3. Comprehensive Robustness Analysis: Systematically evaluates impacts of diagnostic failures and noise
4. High-quality Data: Based on expert-annotated high-quality training data
5. Good Interpretability: Model demonstrates excellent physical interpretability

Weaknesses

1. Single-device Limitation: Verified only on DIII-D; generalization capability requires further validation
2. Auxiliary Diagnostic Dependency: Although primarily using ECE, still requires magnetic diagnostics and density measurements
3. Boundary Conditions: Performance degrades at high densities and when boundary reconstruction fails
4. Dynamic Response: Capability to respond to rapid plasma behavior changes (e.g., ELMs) remains unclear

Impact

1. Field Contribution: Provides practical technical pathway for FPP plasma control
2. Methodological Value: Demonstrates how to design effective ML under constraints
3. Engineering Guidance: Provides reference for future fusion device diagnostic system design

Applicable Scenarios

• Real-time plasma state monitoring in fusion power plants
• Plasma diagnostics in resource-constrained environments
• Plasma control systems requiring high robustness
• Diagnostic system design in other extreme environments

References

The paper cites 33 important references covering ECE diagnostic technology, fusion plasma physics, and machine learning applications in fusion, providing a solid theoretical foundation for the research.

Overall Assessment: This is a high-quality applied physics paper that tightly integrates machine learning technology with fusion plasma physics, proposing innovative solutions to practical engineering requirements. The paper demonstrates excellence in technical innovation, experimental validation, and engineering applicability, making important contributions to plasma control technology development for fusion power plants.