Neural Networks for Predicting Heat Transport in Tokamak Plasmas

Over the past 50 years, there has been remarkable progress in understanding the physics underlying transport phenomena in tokamak plasmas. However, there has yet to be a model  developed that can reliably predict experimental observations for all plasma conditions across the whole plasma radius and, to date, the most complete simulations still pose a challenge to supercomputer resources and compare favorably with experiments only for a limited range of plasma parameters and over a small portion of plasma. This poster presents the idea of using artificial neural networks as a means to perform a non-linear multivariate regression of local transport fluxes as a function of a set of local  dimensionless plasma parameters; compared to existing empirical and theoretical models, the  neural network approach does not make any simplifying assumption about the functional form of the underlying transport physics. For our analysis, three multi-layer, feed-forward, back- propagation neural networks are built and trained using data that is taken from the experimental  databases for the DIII-D, TFTR, and JET tokamaks, and we assume transport to be a local phenomenon, meaning that fluxes at one radial location only depend on plasma parameters at  that same radial location. It is observed that given similar parameters that the most sophisticated  models use, the neural network model is able to predict the heat transport profiles observed in  experiments in DIII-D[1], TFTR, and JET across their entire plasma radii many orders of magnitude more quickly and accurately than any existing methods. Consistent results have been  obtained over a broad spectrum of plasma configurations, including low confinement modes, which the state of-the-art, first-principle, numerical methods are unable to reproduce. A qualitative and quantitative comparison of the performance between the standard methods for predicting heat transport in tokamak plasmas (i.e. first-principle, numerical  methods) and the performance of the neural network approach is presented. Additionally, the  advantages and disadvantages of the neural network are discussed, as well as the potential future applications of our research findings. Future directions for our research are also presented, and we discuss the possibility of determining the relative importance of each of the input parameters  by means of applying a greedy algorithm towards the optimization of the topologies of the neural networks. [1] O. Meneghini, et al., Phys. Plasmas 21 (2014) 060702