Speaker
Description
A. Ho1, L. Zanisi2, B. de Leeuw3, V. Galvan1, P. Rodriguez-Fernandez1, N. T. Howard1
1MIT Plasma Science and Fusion Center, Cambridge, MA, USA
2Culham Centre for Fusion Energy – United Kingdom Atomic Energy Authority, Abingdon, UK
3Radboud University, Nijmegen, Netherlands
This study applies uncertainty-aware neural network architectures in combination with active learning (AL) techniques [L. Zanisi et al., NF 2024] to construct efficient datasets for data-driven surrogate model generation including the simulator in-the-loop. This was applied to the tokamak plasma turbulent transport problem, specifically the QuaLiKiz code [J. Citrin et al., PPCF 2017], generating surrogates aimed at accelerating profile predictions in transport solvers such as PORTALS [P. Rodriguez-Fernandez et al., NF 2024]. This exercise focuses on small datasets as a proxy for using more expensive gyrokinetic codes, e.g. CGYRO [J. Candy et al., Jour. Comp. Sci. 2016], which can be bootstrapped by leveraging gyrokinetic simulation databases. A combination of classifier and regressor model was trained for all turbulent modes (ITG, TEM, ETG) and all transport fluxes provided by QuaLiKiz.
Starting with an initial training data set of 102 points, 45 iterations were performed resulting in a final set of 104 and models with a F1 classification performance of ~0.8 and a R2 regression performance of ~0.8 on an independent test set across all outputs. Additionally, the overall technique is generalizable to create surrogate models beyond the primary domain being studied. This was demonstrated by applying this pipeline on the EPED pedestal stability model [P. Snyder et al., NF 2011], obtaining a R2 of ~0.85 after 22 iterations with a final data set of 103 points.
This work is supported by Commonwealth Fusion Systems under RPP020 and DOE FES under Award DE-SC0024399
| Country or International Organisation | United States of America |
|---|---|
| Affiliation | MIT |
| Speaker's email address | aaronkho@mit.edu |
