Project: ML Simulation of Nuclear Fusion reactors
Description
Thermonuclear fusion holds the promise of generating clean energy on a large scale. One promising approach for controlled fusion power generation is the tokamak, a torus-shaped device that magnetically confines the fusion plasma in its vessel. Currently, not all physical processes in these plasmas are fully understood, ergo, some phenomena cannot be simulated from first principles. Machine learning could prove a useful tool here, utilizing decades of experimental observations to build data-driven simulation tools. This project concerns the simulation of (measurements of) some key aspects of a tokamak plasma’s time evolution, using deep generative models.
We will work with data from the Tokamak à Configuration Variable (TCV) located at EPFL in Lausanne, Switzerland (Figs. a and b). For this device, prior work concerned deep learning methods for classification purposes, i.e., using diagnostic data (for example Fig. c) to classify the state of the plasma into a so-called "Low" (L), "Dithering" (D) or "High" (H) confinement state (Fig. d) [2, 3]. In this project, the aim is to use a similar dataset of measurements to build a generative model of the observables themselves, rather than a discriminative model of a set of labels. While the latter is of great use for data analysis purposes, it is difficult to study plasma behavior from such a model. Learning a generative model (i.e. a simulator) could potentially help in this aspect, for example by aiding our understanding of how intervening on control parameters affects the behavior of a plasma. In general, ML has been used to assist in simulating fusion plasmas in various subdomains and on various levels of complexity, e.g. [4, 5, 6, 7, 8].
Students interested in this project should have a good understanding of the basics of machine learning and deep learning (and the associated programming skills, ideally Python + ML/DL libraries such as PyTorch). A bonus is experience with generative models, (plasma) physics/fusion, and scientific/numerical simulation.
[1] J. M. Tomczak, J. M. “Deep Generative Modeling”. Springer International Publishing (2022)
[2] F. Matos et al. “Classification of tokamak plasma confinement states with convolutional recurrent neural networks” Nucl. Fusion 60 036022
[3] F. Matos et al. “Plasma confinement mode classification using a sequence-to-sequence neural network with attention” Nucl. Fusion 61 046019
[4] K. L. van de Plassche et al. “Fast modeling of turbulent transport in fusion plasmas using neural networks” Physics of Plasmas 27, 022310
[5] J. Seo et al. “Feedforward beta control in the KSTAR tokamak by deep reinforcement learning” Nucl. Fusion 61 106010
[6] J. Abbate et al. “Data-driven profile prediction for DIII-D” Nucl. Fusion 61 046027
[7] V. Gopakumaret et al. "Fourier neural operator for plasma modelling" arXiv preprint arXiv:2302.06542, 2023
[8] Y. Poels et al. “Fast Dynamic 1D Simulation of Divertor Plasmas with Neural PDE Surrogates” arXiv preprint arXiv:2305.18944, 2023
Details
- Supervisor
-
Vlado Menkovski
- Secondary supervisor
-
Yoeri Poels
- Interested?
- Get in contact