A neural network regression allows to emulate the results of first-principle-based turbulent transport
predictions. Model output databases are compiled with inputs spanning a portion of the experimentally relevant parameter space. The neural network then learns the mapping from code inputs to outputs. The speed of the quasilinear models allows for the production of sufficiently large training sets for this application.
QuaLiKiz was used to create a large database of 3.10^8 flux calculations using 1.3 MCPUh on HPC resources (Edison@NERSC). Embedding known physical constraints in the training of the networks is essential for the surrogate model to perform well in transport predictions. As such, we show the importance of choosing the right cost function and more fundamentally, choosing which target variables the networks have to be trained on. Custom figures of merit and visualisation tools were developed to aid with neural network accuracy verification, such as the data slicer which you can use to visualize QuaLiKiz results over the 9D database. Ongoing work is underway for development of higher input dimension (13D) QuaLiKiz-neural-networks, where due to the curse of dimensionality, the relevant input subspace is constrained by an under-construction multi-machine profile database.