[Paper Review] Deep learning based surrogate model for first-principles global simulations of fusion plasmas

Ge Dong, Xishuo Wei+4 more|arXiv (Cornell University)|Jun 21, 2021

Magnetic confinement fusion research36 references4 citations

TL;DR

This paper presents a deep learning surrogate model (SGTC) trained on first-principles gyrokinetic simulations to predict plasma instability properties—specifically growth rates and mode structures—of current-driven kink modes in DIII-D fusion plasmas. With inference times on the order of milliseconds, SGTC enables real-time plasma control by accelerating simulations that are otherwise too slow for practical use in fusion energy applications.

ABSTRACT

The accurate identification and control of plasma instabilities is important for successful fusion experiments. First-principles simulations which can provide physics based instability information including the growth rate and mode structure are generally not fast enough for real-time applications. In this work, a deep-learning based surrogate model as an instability simulator has been developed and trained in a supervised manner with data from the gyrokinetic toroidal code (GTC) global electromagnetic simulations of the current driven kink instabilities in DIII-D plasmas. The inference time of the surrogate model of GTC (SGTC) is on the order of milliseconds, which fits the requirement of the DIII-D real-time plasma control system (PCS). SGTC demonstrates strong predictive capabilities for the kink mode instability properties including the growth rate and mode structure.

Motivation & Objective

• To develop a fast, accurate surrogate model for first-principles global simulations of fusion plasmas to enable real-time instability prediction.
• To address the computational infeasibility of running full gyrokinetic simulations in real time for plasma control applications.
• To train a deep neural network on data from the GTC code to predict key instability characteristics such as growth rate and mode structure.
• To ensure the surrogate model’s inference speed meets the millisecond-level requirements of the DIII-D real-time plasma control system (PCS).
• To validate the surrogate model’s predictive accuracy against high-fidelity first-principles simulations.

Proposed method

• A supervised deep learning framework is employed, training a neural network on a large dataset of input parameters and corresponding instability outputs from GTC simulations.
• The input features include plasma parameters such as current profile, magnetic field configuration, and density gradients relevant to kink mode instabilities.
• The output layer predicts the growth rate and mode structure of the current-driven kink instability in DIII-D plasmas.
• The model architecture is optimized for speed and accuracy, with residual connections and normalization layers to enhance training stability and inference performance.
• The surrogate model is trained using data from global electromagnetic GTC simulations of DIII-D plasmas under varying operational conditions.
• Inference is performed in milliseconds, making it suitable for integration into real-time control systems.

Experimental results

Research questions

• RQ1 Can a deep learning surrogate model accurately predict the growth rate and mode structure of current-driven kink instabilities in DIII-D plasmas?
• RQ2 Does the surrogate model achieve inference speeds compatible with real-time plasma control systems?
• RQ3 How well does the surrogate model generalize across diverse plasma parameter configurations compared to full first-principles simulations?
• RQ4 What is the predictive accuracy of the surrogate model relative to high-fidelity GTC simulations?
• RQ5 Can the surrogate model be trained effectively on a limited set of high-fidelity simulation data to ensure robustness and transferability?

Key findings

• The SGTC surrogate model achieves inference times on the order of milliseconds, satisfying the real-time requirements of the DIII-D plasma control system.
• The model demonstrates strong predictive capability for both the growth rate and mode structure of current-driven kink instabilities across diverse plasma conditions.
• The surrogate model maintains high accuracy when compared to full GTC simulations, with prediction errors remaining within acceptable thresholds for control applications.
• The deep learning framework successfully learns complex nonlinear relationships between plasma parameters and instability characteristics from first-principles simulation data.
• The trained model generalizes well to unseen plasma configurations, indicating robustness and transferability across operational regimes.
• The integration of SGTC into real-time control pipelines is feasible and represents a significant step toward predictive plasma control in magnetic fusion energy.