Abstract Details
Abstracts
Author: Dmitri M Orlov
Requested Type: Poster
Submitted: 2026-04-26 00:42:53
Co-authors: R Clark, V Glukhov, G Subbotin, M Nurgaliev, A Kachkin, D Sorokin, M Austin, J Chen, L Zeng, TL Rhodes, C. Michoski
Contact Info:
UC San Diego
9500 Gilman Drive, 0417
La Jolla, California 92093
US
Abstract Text:
As fusion research advances toward reactor-scale devices, the transition from diagnostic richness in research tokamaks to diagnostic sufficiency in Fusion Power Plants (FPPs) represents a critical challenge for plasma control and state estimation. Future FPP environments will impose severe constraints on sensors, favoring microwave-based diagnostics like electron cyclotron emission (ECE) and profile reflectometry (PR) that can operate through waveguides outside high-radiation zones. We present a parsimonious and robust machine learning framework designed to identify plasma confinement modes—specifically the bifurcation between low-confinement (L-mode) and high-confinement (H-mode)—using a minimal set of these FPP-relevant signals.
The approach employs physics-informed feature extraction: radial basis functions (RBFs) are utilized to interpolate ECE electron temperature profiles, while third-order polynomial splines handle the specific challenges of reflectometer density profiles, such as core-access limitations during high-performance discharges. These features are processed through gradient boosting classifiers, achieving test accuracies of 96% for ECE and 97% for PR alone. To enhance reliability, an ensemble model was developed that integrates both diagnostics via a weighted average probability informed by k-means clustering for uncertainty quantification. This ensemble achieves 99.2% test accuracy and remains resilient to significant signal noise and sensor failure.
Looking forward, this work outlines an "Agentic AI" framework for real-time (sub-10 ms) plasma state reconstruction. By utilizing multi-fidelity surrogate models trained on high-fidelity TRANSP interpretive databases, we demonstrate the potential to infer latent plasma parameters—such as equilibrium structure and plasma beta—from sparse, noisy measurements. This methodology provides a scalable pathway for autonomous plasma control and machine protection in next-generation fusion devices.
Characterization: 7.0
Comments: