April 4-6

Abstract Details

files Add files

status:file name:submitted:by:
approvedzhu_sherwood_abs-1.pdf2022-03-11 14:51:46Ben Zhu


Author: Ben Zhu
Requested Type: Poster
Submitted: 2022-03-11 14:50:02

Co-authors: M.Zhao, H.Bhatia, B.Meyer, T.Rognlien, T.Bremer, X.Q.Xu

Contact Info:
Lawrence Livermore National Laboratory
7000 East Avenue, L-440
Livermore, CA   94550

Abstract Text:
Divertor power exhaust handling is a key component of the success for future tokamak reactors. So far, the most successful power reduction is achieved by detachment. We have developed a physics model-based detachment prediction with machine learning techniques by finding robust and accurate projections among three distinct descriptions of steady-state plasma states — model inputs (i.e., engineering parameters such as heating power, puffing rate/upstream density), omniscient diagnostics (ODs, such as synthetic Langmuir probe, Thomson scattering, radiation measurements), and latent space that identified by an autoencoder when compressing the ODs. As a starting point, we use highly efficient 1D UEDGE model which contains the crucial physics ingredi- ents of detachment (e.g., ionization and recombination of plasma and neutrals) to simulate the plasma and neutrals along the open magnetic field lines in the Scrape-Off Layer. Over 300,000 simulations with varying model inputs (e.g., different upstream density, power, carbon fraction and divertor leg, or, magnetic connec- tion length) are performed to generate the training database; and we find that a 6-dimensional bottleneck layer (6D) in latent space is good enough to closely yield a match for our true system in configuration space (i.e., the ODs such as upstream temperature, divertor target density, temperature and saturation current, as well as radiation profile). Based on this finding, forward surrogate models are trained to make predictions into the bottleneck layer from model inputs; then the trained decoder is used to reconstruct back the ODs to the configuration space. Current proof-of-principle model performs similarly to the analytical two-point model but with local flux-limited thermal transport and features additional detachment front prediction. Moreover, this approach can be easily extended to incorporate richer physics in a more realistic experiment setting once trained upon a higher fidelity dataset.