
Machine learning for classification, dimension reduction, clustering, and feature importance analysis of QH-mode plasma regimes
Matthew Waller, Zeyu Li, Rachel A Myers, Craig Michoski, Brian S Sammuli, David Orozco, Tapan Ganatma Nakkina, Xi Chen, Tomas Odstrcil, Andrew Oakleigh Nelson, et al.
Nuclear FusionPlasma & ConfinementControl & DiagnosticsAI, Modeling & Simulation
Edge-localized modes (ELMs) release bursts of energy and particles that threaten plasma-facing components in future tokamak reactors, making ELM-free operating regimes a priority for fusion research. Quiescent H-mode (QH), wide-pedestal QH-mode (WPQH), and broadband QH-mode (BBQH) form a family of ELM-free regimes at DIII-D whose boundaries and controlling parameters remain incompletely understood. This work aims to identify, from data, which measured quantities distinguish the three regimes, using a physics-labeled database of 482 high-confidence segments drawn from 323 DIII-D discharges, whose manual labeling was accelerated by the Data Fusion Labeler (dFL). An XGBoost classifier trained on 1 ms-resolution diagnostic data establishes that the manual labels are learnable: under leave-one-shot-out cross-validation it reproduces them for 87.3% of 1 ms samples (macro-F1 0.82). The same classifier underpins dFL's autolabeler for extending the database across the archive, at an estimated ~50× speedup over fully manual analysis. A supervised feature-importance analysis (classifiers evaluated on held-out discharges, with SHAP attribution), trained only on physics signals not used as labeling criteria, shows that the regimes are also distinguishable from independent physics quantities, and unsupervised dimension-reduction and clustering methods examine the same data without labels. The analysis recovers injected torque as the dominant actuator distinguishing the regimes. Within the rotation and E_r quantities considered, edge E×B shear carries the most information distinguishing the three regimes, while toroidal rotation also contributes strongly; this is a statistical association for this model and dataset, consistent with the critical-shear picture of edge harmonic oscillation onset. Where the supervised and unsupervised rankings appear to disagree, the disagreement reflects the correlation structure among related features rather than a substantive physics dispute. Together, these results identify key physical quantities associated with QH-mode regime boundaries and provide a reusable, interpretable workflow for machine-learning-based analysis of labeled databases in fusion research.


