Physics-Informed Visual MARFE Prediction on the HL-3 Tokamak

TL;DR

This work tackles the critical challenge of predicting MARFE onset and enabling disruption mitigation on the HL-3 tokamak. It introduces a physics-informed pipeline that refines noisy camera-based MARFE labels using a weighted EM algorithm with a physics prior, and a continuous-time Neural ODE predictor with a physics gate to forecast short-horizon MARFE worsening. The framework demonstrates high predictive performance (AUC ≈ 0.959 for 40 ms ahead) and successful real-time deployment at a 1 ms cycle, validating its viability for proactive control and mitigation strategies. The combination of physically grounded label refinement and continuous-time dynamics provides a robust, physics-consistent indicator poised to inform next-generation devices such as ITER.

Abstract

The Multifaceted Asymmetric Radiation From the Edge (MARFE) is a critical plasma instability that often precedes density-limit disruptions in tokamaks, posing a significant risk to machine integrity and operational efficiency. Early and reliable alert of MARFE formation is therefore essential for developing effective disruption mitigation strategies, particularly for next-generation devices like ITER. This paper presents a novel, physics-informed indicator for early MARFE prediction and disruption warning developed for the HL-3 tokamak. Our framework integrates two core innovations: (1) a high-fidelity label refinement pipeline that employs a physics-scored, weighted Expectation-Maximization (EM) algorithm to systematically correct noise and artifacts in raw visual data from cameras, and (2) a continuous-time, physics-constrained Neural Ordinary Differential Equation (Neural ODE) model that predicts the short-horizon ``worsening" of a MARFE. By conditioning the model's dynamics on key plasma parameters such as normalized density ($f_G$, derived from core electron density) and core electron temperature ($T_e$), the predictor achieves superior performance in the low-false-alarm regime crucial for control. On a large experimental dataset from HL-3, our model demonstrates high predictive accuracy, achieving an Area Under the Curve (AUC) of 0.969 for 40ms-ahead prediction. The indicator has been successfully deployed for real-time operation with updates every 1 ms. This work lays a very foundation for future proactive MARFE mitigation.

Figures (10)

• Figure 1: The data processing pipeline for generating cleaned visual features. Initial area features (i.e., $m_U$, $m_M$ and $m_L$), which are extracted from raw images, are used to derive an initial binary label $y_{\text{init}}$. An EM algorithm then produces a refined label $\hat{y}_i$, which in turn is used to clean the initial features to produce the final model inputs ($m'_U, m'_M, m'_L$).
• Figure 2: The image processing pipeline for preliminary MARFE feature extraction and its correlation with a MARFE-induced disruption on shot #11595. (Top) The temporal evolution leading to a disruption: (a) a stable plasma state before the MARFE at $t=900$ ms, (b) the onset of the MARFE at the high-field side at $t=1120$ ms, and (c) the subsequent plasma disruption at $t=1220$ ms, characterized by intense, widespread light emission. (Middle) The sequential processing steps applied to the onset frame (b): (d) the ROI-masked grayscale image $M_{\text{RoI}}$ isolates the region of interest, (e) the image is binarized to identify MARFE candidates, and (f) a morphological opening produces the final, denoised feature map. (Bottom) A time-series comparison of the raw and refined MARFE area signals.
• Figure 3: Probability density distributions for key physical parameters ($n_e$, $T_e$, $f_G$) separated by MARFE and non-MARFE states. The vertical dashed lines indicate the thresholds determined through data-driven analysis (ROC curves and distribution percentiles), which are used to calculate the physics-based prior scores.
• Figure 4: Schematic illustration of the physics-informed label refinement process using a weighted EM algorithm. The process begins with initial noisy MARFE labels and the associated physical parameters ($\mathbf{x}_i = [n_e, T_e, f_G, t]_i$). A physics prior score, $s_i$, is constructed from these parameters to quantify the propensity for MARFE formation. The physical features are modeled as a two-component Gaussian Mixture Model (GMM), representing the MARFE (pos) and non-MARFE (neg) states. The algorithm then iteratively refines the labels: in the E-step, the physics prior $s_i$ is critically used as a sample-specific prior to calculate the posterior probability, or responsibility, $\gamma(z_{i,\text{pos}})$. In the M-step, these responsibilities are used to update the GMM parameters ($\boldsymbol{\mu}, \boldsymbol{\Sigma}$). After convergence, the final posterior probabilities are clipped with a threshold to produce the refined, clean labels.
• Figure 5: The overall architecture of the physics-informed MARFE prediction framework. The Neural ODE Model, which encodes historical time series data, evolves the system's latent state continuously with a physics-gated Neural ODE, and decodes the future state to predict MARFE worsening.