Fast modelling of turbulent transport in fusion plasmas using neural networks

Fast modelling of turbulent transport in fusion plasmas using neural networks

Fast modelling of turbulent transport in fusion plasmas using neural networks 150 150 amit
UKAEA-CCFE-PR(20)122

Fast modelling of turbulent transport in fusion plasmas using neural networks

We present an ultrafast neural network (NN) turbulent tokamak transport model, QLKNN, for heat and particle fluxes. QLKNN is a surrogate model based on a database of 3 · 108 flux calculations of the quasilinear gyrokinetic trans- port model QuaLiKiz. To ensure accurate reproduction of the underlying model, we include known features of the physical system by choosing specific training targets and using a customized cost function in our training pipeline. We have coupled QLKNN to the tokamak modelling framework JINTRAC and rapid control-oriented tokamak trans- port solver RAPTOR. We demonstrate and validate the coupled frameworks through application to three JET shots covering a representative spread of H-mode operating space, predicting turbulent transport in the plasma core region (0.2 < ρN,tor < 0.85). QLKNN is able to accurately reproduce QuaLiKiz-predicted kinetic profiles (Ti,e and ne ) but or- ders of magnitude faster, from 7 days on 16 cores (JINTRAC-QuaLiKiz) to 20 minutes on 2 cores (JINTRAC-QLKNN) and 90 seconds on 1 core (RAPTOR-QLKNN). The discrepancy between QLKNN and QuaLiKiz is only on the or- der 1%-10% in rotationless cases. The impact of rotation on turbulent fluxes is included in QLKNN through a new flux scaling rule in postprocessing, based on a set of linear gyrokinetic simulations. This difference from the native QuaLiKiz rotation rule results in slightly larger (3%-15%) differences in the final kinetic profiles for cases including rotation. Dynamic behaviour is also well captured by QLKNN, with differences of only 4%-10% compared to full Qua- LiKiz observed at mid-radius, for a study of density buildup following the LH transition. Deployment of neural network surrogate models in multi-physics integrated tokamak modelling is a promising route towards enabling accurate and fast tokamak scenario optimization, Uncertainty Quantification, and control-oriented applications.

Collection:
Journals
Journal:
Physics of Plasmas
Publisher:
AIP (American Institute of Physics)
Published date:
02/11/2020