Towards fast surrogate models for interpolation of tokamak edge plasmas

One of the major design limitations for tokamak fusion reactors is the heat load that can be sustained by the materials at the divertor target. Developing a full understanding of how machine or operation parameters affect the conditions at the divertor requires an enormous number of simulations. A promising approach to circumvent this is to use machine learning models trained on simulation data as surrogate models. Once trained such surrogate models can make fast predictions for any scenario in the design parameter space. In future such simulation based surrogate models could be used in system codes for rapid design studies of future fusion power plants. This work presents the first steps towards the development of such surrogate models for plasma exhaust and the datasets required for their training. Machine learning models like neural networks usually require several thousand data points for training, but the exact amount of data required varies from case to case. Due to the long runtimes of simulations we aim at finding the minimal amount of training data required. A preliminary dataset based on SOLPS-ITER simulations with varying tokamak design parameters, including the major radius, magnetic field strength and neutral density is constructed. To be able to generate more training data within reasonable computation time the simulations in the dataset use fluid neutral simulations and no fluid drift effects. The dataset is used to train a simple neural network and Gradient Boosted Regression Trees and test how the performance depends on the number of training simulations.

Automated summary

One major limitation for tokamak fusion reactors is the heat load on the divertor target. Machine learning models trained on simulation data can provide fast predictions for various scenarios in the design parameter space, and thus be used in system codes for rapid design studies of fusion power plants. This work aims to develop surrogate models for plasma exhaust using a preliminary dataset and train machine learning models to test their performance.