A generative artificial intelligence surrogate model of plasma turbulence
ORAL
Abstract
Turbulent transport plays a key role in confinement degradation limiting the performance of current and future fusion devices. Modelling turbulent transport requires long- time simulations which are limited by the computational resources available. One way to overcome this shortcoming is by using surrogate models that are computationally cheaper to evaluate. In this presentation we apply generative artificial intelligence (AI) methods to construct a surrogate model of plasma edge turbulence described by the HW (Hasegawa-Wakatani) model and use the model to perform fast, long-time turbulent transport computations.
The proposed GAIT (Generative Artificial Intelligence Turbulence) model is based on the combination of a convolutional variational autoencoder and a deep neural network (DNN). A convolutional network is used to encode snapshots of computed HW turbulence states into a reduced latent space, and a DNN is trained to reproduce the time evolution of turbulence in the latent space. Once the autoencoder is trained, new turbulence states are obtained by decoding the latent space dynamics generated by the DNN.
To evaluate the model we use Eulerian and Lagrangian metrics. Good agreement is found between the GAIT and the HW models in the spatial and temporal Eulerian turbulence Fourier spectra. In the Lagrangian setting, the statistical moments and probability distribution of particle displacements are compared, and agreement is found in the effective diffusivity in the GAIT and the HW models.
The proposed GAIT (Generative Artificial Intelligence Turbulence) model is based on the combination of a convolutional variational autoencoder and a deep neural network (DNN). A convolutional network is used to encode snapshots of computed HW turbulence states into a reduced latent space, and a DNN is trained to reproduce the time evolution of turbulence in the latent space. Once the autoencoder is trained, new turbulence states are obtained by decoding the latent space dynamics generated by the DNN.
To evaluate the model we use Eulerian and Lagrangian metrics. Good agreement is found between the GAIT and the HW models in the spatial and temporal Eulerian turbulence Fourier spectra. In the Lagrangian setting, the statistical moments and probability distribution of particle displacements are compared, and agreement is found in the effective diffusivity in the GAIT and the HW models.
*This work has received financial support from the AIM4EP project (ANR-21-CE30-0018), funded by the French National Research Agency (ANR), and from the Oak Ridge National Laboratory, managed by UTBattelle, LLC, for the US Department of Energy under Contract No. DE-AC05-00OR22725.All the simulations and training of neural networks reported here were performed on HPC resources of IDRIS under the allocations 2021-A0100512455, 2022-AD010512455R1 and 2023-A0140514165 made by GENCI.
–
Publication: Clavier, B., et al. "A generative machine learning surrogate model of plasma turbulence." arXiv preprint arXiv:2405.13232 (2024).
Presenters
-
Benoît Clavier