arXiv subjects

Wulyu Zhong

Publications and source records attributed to Wulyu Zhong.

7 recordsLinked to original sources

arXiv2026

Physics-Gated Visual Prediction of MARFE on the HL-3 Tokamak

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. We develop a physics-gated, continuous MARFE monitor for the HL-3 tokamak that outputs a per-frame intensity probability every $2$\,ms, which can potentially be used by the shape-target controller in the plasma control system. Our framework integrates two core innovations: (1) a physics-scored, weighted Expectation-Maximization (EM) pipeline that refines noisy visual labels using $(n_e, T_e, f_G, t)$ as a Bayesian prior, and (2) a continuous-time, physics-gated Neural Ordinary Differential Equation (Neural ODE) backbone whose dynamics are modulated by a sigmoid gate on $f_G$ and $T_e$. Meanwhile, the Neural ODE adopts a $40$\,ms forward forecasting horizon to accommodate the actuator-response budget. On a frozen $140$-shot held-out test set, the proposed method yields a median label-aligned lead time of $+36$\,ms, close to this design horizon. Against a Bi-LSTM baseline trained under the matched protocol, the proposed Neural ODE attains Area Under the Curve (AUC) $=0.981$ and sample-level $F_1=0.840$, compared with AUC $=0.960$ and sample-level $F_1=0.779$ for the baseline. The deployed inference service runs within a $1$-ms control-cycle budget, while new diagnostic samples are generated at the $2$-ms frame cadence.

physics.plasm-ph↗

arXiv2025

Plasma Shape Control via Zero-shot Generative Reinforcement Learning

Traditional PID controllers have limited adaptability for plasma shape control, and task-specific reinforcement learning (RL) methods suffer from limited generalization and the need for repetitive retraining. To overcome these challenges, this paper proposes a novel framework for developing a versatile, zero-shot control policy from a large-scale offline dataset of historical PID-controlled discharges. Our approach synergistically combines Generative Adversarial Imitation Learning (GAIL) with Hilbert space representation learning to achieve dual objectives: mimicking the stable operational style of the PID data and constructing a geometrically structured latent space for efficient, goal-directed control. The resulting foundation policy can be deployed for diverse trajectory tracking tasks in a zero-shot manner without any task-specific fine-tuning. Evaluations on the HL-3 tokamak simulator demonstrate that the policy excels at precisely and stably tracking reference trajectories for key shape parameters across a range of plasma scenarios. This work presents a viable pathway toward developing highly flexible and data-efficient intelligent control systems for future fusion reactors.

physics.plasm-ph↗

arXiv2025

FusionMAE: large-scale pretrained model to optimize and simplify diagnostic and control of fusion plasma

In magnetically confined fusion device, the complex, multiscale, and nonlinear dynamics of plasmas necessitate the integration of extensive diagnostic systems to effectively monitor and control plasma behaviour. The complexity and uncertainty arising from these extensive systems and their tangled interrelations has long posed a significant obstacle to the acceleration of fusion energy development. In this work, a large-scale model, fusion masked auto-encoder (FusionMAE) is pre-trained to compress the information from 88 diagnostic signals into a concrete embedding, to provide a unified interface between diagnostic systems and control actuators. Two mechanisms are proposed to ensure a meaningful embedding: compression-reduction and missing-signal reconstruction. Upon completion of pre-training, the model acquires the capability for 'virtual backup diagnosis', enabling the inference of missing diagnostic data with 96.7% reliability. Furthermore, the model demonstrates three emergent capabilities: automatic data analysis, universal control-diagnosis interface, and enhancement of control performance on multiple tasks. This work pioneers large-scale AI model integration in fusion energy, demonstrating how pre-trained embeddings can simplify the system interface, reducing necessary diagnostic systems and optimize operation performance for future fusion reactors.

physics.plasm-ph↗

arXiv2025

High-Fidelity Data-Driven Dynamics Model for Reinforcement Learning-based Control in HL-3 Tokamak

The success of reinforcement learning (RL)-based control in tokamaks, an emerging technique for controlled nuclear fusion with improved flexibility, typically requires substantial interaction with a simulator capable of accurately evolving the high-dimensional plasma state. Compared to first-principle-based simulators, whose intense computations lead to sluggish RL training, we devise an effective method to acquire a fully data-driven simulator, by mitigating the arising compounding error issue due to the underlying autoregressive nature. With high accuracy and appealing extrapolation capability, this high-fidelity dynamics model subsequently enables the rapid training of a qualified RL agent to directly generate engineering-reasonable actuator commands, aiming at the desired long-term targets of plasma configuration. Together with a surrogate model for Equilibrium Fitting code based on neural network, named EFITNN, the RL agent successfully maintains a 400-ms, 1 kHz trajectory control with accurate waveform tracking of plasma current and last closed flux surface on the HL-3 tokamak. Furthermore, it also demonstrates the feasibility of zero-shot adaptation to changed triangularity targets, confirming the robustness of the developed data-driven dynamics model. Our work underscores the advantage of fully data-driven dynamics models in yielding RL-based trajectory control policies at a sufficiently fast pace, an anticipated engineering requirement in daily discharge practices for the upcoming ITER device.

physics.plasm-ph↗

arXiv2024

Adapted Swin Transformer-based Real-Time Plasma Shape Detection and Control in HL-3

In the field of magnetic confinement plasma control, the accurate feedback of plasma position and shape primarily relies on calculations derived from magnetic measurements through equilibrium reconstruction or matrix mapping method. However, under harsh conditions like high-energy neutron radiation and elevated temperatures, the installation of magnetic probes within the device becomes challenging. Relying solely on external magnetic probes can compromise the precision of EFIT in determining the plasma shape. To tackle this issue, we introduce a real-time, non-magnetic measurement method on the HL-3 tokamak, which diagnoses the plasma position and shape via imaging. Particularly, we put forward an adapted Swin Transformer model, the Poolformer Swin Transformer (PST), to accurately and fastly interpret the plasma shape from the Charge-Coupled Device Camera (CCD) images. By adopting multi-task learning and knowledge distillation techniques, the model is capable of robustly detecting six shape parameters under disruptive conditions such as a divertor shape and gas injection, circumventing global brightness changes and cumbersome manual labeling. Specifically, the well-trained PST model capably infers R and Z within the mean average error below 1.1 cm and 1.8 cm, respectively, while requiring less than 2 ms for end-to-end feedback, an 80 improvement over the smallest Swin Transformer model, laying the foundation for real-time control. Finally, we deploy the PST model in the Plasma Control System (PCS) using TensorRT, and achieve 500 ms stable PID feedback control based on the PST-computed horizontal displacement information. In conclusion, this research opens up new avenues for the practical application of image-computing plasma shape diagnostic methods in the realm of real-time feedback control.

physics.plasm-ph↗

arXiv2024

Real-time equilibrium reconstruction by neural network based on HL-3 tokamak

A neural network model, EFITNN, has been developed capable of real-time magnetic equilibrium reconstruction based on HL-3 tokamak magnetic measurement signals. The model processes inputs from 68 channels of magnetic measurement data gathered from 1159 HL-3 experimental discharges, including plasma current, loop voltage, and the poloidal magnetic fields measured by equilibrium probes. The outputs of the model feature eight key plasma parameters, alongside high-resolution ($129\times129$) reconstructions of the toroidal current density $J_{\text P}$ and poloidal magnetic flux profiles $Ψ_{rz}$. Moreover, the network's architecture employs a multi-task learning structure, which enables the sharing of weights and mutual correction among different outputs, and lead to increase the model's accuracy by up to 32%. The performance of EFITNN demonstrates remarkable consistency with the offline EFIT, achieving average $R^2 = 0.941, 0.997$ and $0.959$ for eight plasma parameters, $Ψ_{rz}$ and $J_{\text P}$, respectively. The model's robust generalization capabilities are particularly evident in its successful predictions of quasi-snowflake (QSF) divertor configurations and its adept handling of data from shot numbers or plasma current intervals not previously encountered during training. Compared to numerical methods, EFITNN significantly enhances computational efficiency with average computation time ranging from 0.08ms to 0.45ms, indicating its potential utility in real-time isoflux control and plasma profile management.

physics.plasm-ph↗

arXiv2024

Identifying L-H transition in HL-2A through deep learning

During the operation of tokamak devices, addressing the thermal load issues caused by Edge Localized Modes (ELMs) eruption is crucial. Ideally, mitigation and suppression measures for ELMs should be promptly initiated as soon as the first low-to-high confinement (L-H) transition occurs, which necessitates the real-time monitoring and accurate identification of the L-H transition process. Motivated by this, and by recent deep learning boom, we propose a deep learning-based L-H transition identification algorithm on HL-2A tokamak. In this work, we have constructed a neural network comprising layers of Residual Long Short-Term Memory (LSTM) and Temporal Convolutional Network (TCN). Unlike previous work based on recognition for ELMs by slice, this method implements recognition on L-H transition process before the first ELMs crash. Therefore the mitigation techniques can be triggered in time to suppress the initial ELMs bursts. In order to further explain the effectiveness of the algorithm, we developed a series of evaluation indicators by shots, and the results show that this algorithm can provide necessary reference for the mitigation and suppression system.

physics.plasm-ph↗