[論文レビュー] Plasma Confinement State Classification in Fusion Power Plants: Profile Reflectometer and Ensemble Diagnostics

Randall Clark, Vacslav Glukhov+5 more|arXiv (Cornell University)|Feb 2, 2026

Magnetic confinement fusion research被引用数 0

ひとこと要約

tldr: The paper develops a Profile Reflectometer (PR) based plasma confinement mode classifier and an ensemble model combining PR with an existing ECE-based detector to identify H-mode in fusion plasmas with high accuracy, aiming for reactor-relevant diagnostics in Fusion Power Plants. The PR model achieves 97% test accuracy, and the ensemble achieves 99.2% test accuracy.

ABSTRACT

As Fusion Pilot Plants (FPPs) are increasingly viewed as within reach, many engineering challenges remain. Not many diagnostics are expected to be available in a reactor environment. Survivability, maintainability, and limited port space substantially restrict the number of FPP-relevant diagnostics. One remaining challenge is developing tools and devices to extract plasma state information necessary for controlling an FPP from a limited subset of diagnostics. This work is part of an overarching project to address this challenge. The specific diagnostic subset to be used in FPPs is still under debate. We take the approach of developing machine-learning-based tools for different significant plasma state parameters, using already known FPP-viable diagnostics. Previously we developed a plasma confinement mode classifier utilizing the Electron Cyclotron Emission (ECE) diagnostic. Here, we expand on this by developing a Profile Reflectometer (PR) based classifier with 97\% test accuracy, and an ensemble model that combines the ECE and PR models into a single model, achieving 99\% test accuracy.

研究の動機と目的

• Motivate the need for robust, reactor-relevant diagnostics under limited diagnostics port space in Fusion Power Plants (FPPs).
• Develop a Profile Reflectometer (PR) based classifier to identify H-mode confinement from limited diagnostics.
• Combine PR with an Electron Cyclotron Emission (ECE) based classifier into an ensemble to improve accuracy and robustness.
• Assess model robustness to future data and potential unseen plasma states.
• Provide a benchmark for PR-based confinement mode classification on DIII-D as a reference for FPP design.

提案手法

• Profile Reflectometer data are processed by fitting 3rd-order splines with 10 knots to generate 10 feature values along the normalized radius (rho* = [0,0.2,0.4,0.6,0.8,0.85,0.9,0.95,1.0,1.1]).
• A Gradient Boosted Classifier (GBC) from sklearn is trained on the spline outputs to perform binary L/H (L-mode vs H-mode) classification.
• Handling incomplete core data by padding with the deepest known value and focusing on edge/pedestal information for H-mode detection.
• An ensemble model combines PR and ECE predictions using a reliability-weighted average with confidence scores derived from a k-means based density-coverage scheme to quantify model confidence.
• Robustness checks include 100 reshuffles of shots for PR, ECE, and ensemble, and a sliding-window temporal test to simulate future data shifts.

Figure 1 : Distribution of the PR data used in training and testing according to the time within the shot, and the $\rho$ range.

実験結果

リサーチクエスチョン

• RQ1 Can a Profile Reflectometer (PR) based approach accurately classify plasma confinement mode (L vs H) using edge/pedestal density information when core data may be incomplete?
• RQ2 Does combining PR with the existing ECE-based classifier improve confinement mode identification accuracy and robustness?
• RQ3 How does the ensemble handle uncertainties and partial feature space coverage via confidence weighting?
• RQ4 What is the ensemble’s performance under temporal data shifts that simulate future reactor conditions?
• RQ5 What are the practical implications for deploying FPP-relevant diagnostics with constrained diagnostics ports?

主な発見

• PR-based confinement mode classifier achieves 97% test accuracy on 8102 data points from 260 shots with 80/20 train/test split.
• Edge/pedestal region dominates the PR model inputs; Shapley analysis confirms edge data are most informative for H-mode detection.
• Ensemble model combining PR and ECE achieves 99.2% test accuracy with 0.79% standard deviation across 100 reshuffles.
• Sliding-window tests show robustness to future data with ECE 91.1%, PR 93.1%, Ensemble 96.0% average accuracy, indicating some degradation relative to randomized tests but still high performance.
• The study demonstrates value in attaching model confidence to each diagnostic to address complementary blind spots between PR and ECE.

Figure 2 : A t-SNE heuristic visualization of PR labeled data. Apart from a few outliers at the boundary between L- and H-mode, the hand-labeled data points separate cleanly.