Summary of the Project
This project aims to develop a comprehensive suite of machine learning tools for predicting turbulent transport in tokamak devices. To achieve this goal, we will employ the newly developed T3ST code—a highly parallelized Fortran-based numerical solver designed to follow the dynamics of charged particles in tokamak-like electromagnetic environments. Turbulence is incorporated in an ad-hoc manner through synthetically generated random fields formulated in field-aligned coordinates. Consequently, T3ST is deeply rooted in the statistical paradigm of turbulence representation and is capable of computing transport coefficients through Lagrangian averaging techniques.
Within the scope of this project, T3ST will be used to evaluate transport coefficients across a wide range of scenarios defined by both equilibrium and turbulent properties. The relevant parameter space is expected to exceed 10–15 dimensions, and between 10⁵ and 10⁶ simulations will be performed. The resulting transport coefficients will form a large-scale numerical database that captures the dependence of turbulent transport on the chosen parameters.
In the final phase of the project, this database will serve as the foundation for training and validating machine learning models—primarily feed-forward deep neural networks supported by regression-based approaches. These models will ultimately provide fast, accurate predictions of turbulent transport, offering a powerful complement to traditional first-principles simulations.