Surrogate model of transport quantities of gyrokinetic simulations on HL-2A

Abstract

Artificial intelligence techniques, particularly machine learning, are increasingly being employed to reduce the computational cost of high-fidelity fusion simulations. In this work, nonlinear gyrokinetic simulations are performed to systematically scan the plasma gradient parameter space relevant to HL-2A discharge #27055 plasmas, generating a database of electrostatic drift-wave turbulence for the development of transport surrogate models. A local surrogate model for turbulent transport is first constructed using a feedforward neural network (FNN) trained on 4267 nonlinear local simulation data points, achieving coefficients of determination of up to R2≃0.9 for key transport coefficients. Building upon this, a second-stage global surrogate model based on support vector regression (SVR) is developed to map local transport predictions to global turbulent transport, trained on 40 global gyrokinetic simulation cases. The resulting global surrogate model is subsequently coupled to an reduced transport solver, which successfully reproduces the steady-state plasma profiles in good agreement with experimental observations. This study provides a practical workflow for predicting global turbulent transport using local surrogate models and machine learning techniques. The proposed approach offers an accurate and computationally efficient strategy for integrated transport simulations, with potential applications in scenario development and control-oriented real-time simulations for present and future fusion devices.