Comparison of neural network and gradient-boosted decision tree surrogates of linear gyrokinetic simulations for turbulent transport modeling in tokamak plasmas

A principal computational bottleneck in the integrated modeling of tokamak plasmas is the evaluation of turbulent transport fluxes, with even reduced quasi-linear gyrokinetic models remaining orders of magnitudes too slow for large scale predictive simulations and real-time applications. Machine learning surrogate models offer a route to accelerate integrated modeling by replacing existing turbulent transport models with fast predictive surrogates. We present a systematic comparison of two widely used algorithms: neural networks (NNs), and gradient-boosted decision trees (XGBoost) to predict linear stability and unstable-mode growth rates, trained on linear QuaLiKiz simulations spanning a 22-dimensional input space covering the experimental domain of the JET tokamak. Model performance is assessed in a heavily optimized comparison study using 105 training points and a scaling study covering 102 to 107 training points. NNs across both studies were significantly more sensitive requiring more tuning and diagnostic effort to match the predictive capability of the XGBoost models. Both models achieved comparable performance at 105 data points when heavily optimized but XGBoost pulls ahead in the scaling study with reduced optimization and demonstrates improved robustness of model hyperparameters. XGBoost additionally delivers ∼50× faster inference and ∼35× faster training and hyperparameter-optimization times across the board. The reduced training times and increased robustness of XGBoost are beneficial for faster and cheaper model development, especially in active learning-based pipelines, while its faster inference speeds are ideal for rapid integrated modeling applications. The primary drawback to XGBoost is its step-function predictions with discontinuous derivatives, which may be problematic in certain applications.