AbstractIntegrated simulation for output profiles of a nuclear fusion reactor is generally too slow, when using first-principle-based modeling, to be useful for real-time plasma dynamics control purposes. Neural networks hold the promise of delivering surrogate models that are sufficiently fast to operate in real-time. Here we present the results of using a deep generative model as a neural network alternative for integrated simulation suites. First, by training a conditional variational auto-encoder on a simple synthetic dataset, the quality of generated data is investigated, yielding high-quality predictions with a relative error of only 1.38% on a testing set. We also examine the role that the latent space plays in this generation, andfind it to effectively take the role of all relevant parameters that are the source of variation between the samples. Then, we introduce the more powerful regressive disentanglement variational auto-encoder (ReD-VAE), which is applied to a much larger and complex experimental dataset. By introducing an additional loss term, the quality of generated data is
improved significantly with respect to a normal VAE. This leads to excellent performance in the prediction of known scaling laws, which are used to validate the correctness of model predictions. Other usages of ReD-VAE are also presented, including statistical inference of experimental parameters with good accuracy. A proof-of-principle is provided of using the latent space for unsupervised hidden parameter discovery. Plasma mode is expected to be the main hidden parameter (to the model), which can be reconstructed with 86% accuracy in an unsupervised fashion, showing that this model can potentially be used to gain further insights into what other hidden parameters influence tokamak performance.
|Date of Award||2 Mar 2021|
|Supervisor||J.M.V.A. (Vianney) Koelman (Supervisor 1), V. Menkovski (Supervisor 1), Jonathan Citrin (Supervisor 1) & Karel L. van de Plassche (Supervisor 1)|
- nuclear fusion
- Deep Neural Network