Transport model surrogates using stochastic variational Gaussian processes with deep kernels

A surrogate model for turbulent transport models is proposed based on Gaussian process (GP) regression. To ensure applicability to medium- and large-scale datasets, the model incorporates the stochastic variational Gaussian process, a scalable GP framework. Expressive deep multi-layer kernels, constructed hierarchically in a layer-by-layer manner using an iterative formula, are employed to capture the complex input–output relationship characteristics of turbulent transport models. Integrated into the steady-state transport code GOTRESS, the GP-based surrogate models successfully reproduce the temperature profiles predicted by the original turbulent transport model TGLF. It is confirmed that the GP-based surrogate model remains effective even when trained on a very limited number of training data points.