Development of a Physics-Constrained Machine Learning Model for Generalized Quasi-Linear Diffusion Coefficient Prediction in Lower Hybrid Current Drive Scenarios

Abstract

In this thesis, machine learning models for predicting the quasi-linear diffusion coefficient (D_QL) are explored, compared, and evaluated. A novel machine learning framework with physical constraints specifically designed to predict D_QL is proposed. The proposed D_QL prediction model successfully reproduces the ground truth D_QL obtained from GENRAYCQL3D simulations. Furthermore, when the predicted DQL replaces the conventionally computed D_QL in GENRAY-CQL3D, the resulting radial current drive profiles exhibit consistent behavior with those from the original simulation, while achieving a substantial reduction in computational time. These results demonstrate that the developed surrogate model not only accurately captures key wave–particle interaction features but also significantly accelerates simulation efficiency while maintaining accuracy. Accurate prediction of the quasi-linear diffusion coefficient is regarded as crucial for describing wave–particle interactions and non-inductive current drive in tokamak plasmas. However, conventional methods for evaluating D_QL rely on computationally intensive wave propagation and Fokker–Planck simulations. To address this challenge, a machine learning-based surrogate model capable of predicting D_QL both rapidly and accurately is developed. Initially, to predict D_QL, dimensionality reduction methods combined with neural networks are explored. Although these approaches enable improved computational efficiency, they struggle to fully preserve fine-scale details essential for accurate current profile prediction. To overcome these limitations, a hybrid architecture combining a U-Net and an autoregressive neural network (ARNN) is introduced. This hybrid architecture effectively captures the physics inherent in D_QL, while incorporating physical constraints via the Potential Power Deposition (PPD) method. Overall, the final U-Net–ARNN framework consistently outperforms other approaches, delivering precise wave–particle interaction modeling with greatly enhanced computational efficiency.