Abstract Details
Abstracts
Author: Kevin S. Gill
Requested Type: Consider for Invited
Submitted: 2026-03-10 17:43:46
Co-authors: I. G. Farcas, S. Glas, B. J. Faber, A. Wright
Contact Info:
University of Wisconsin-Madison
1500 Engineering Dr
Madison, WI 53706
United States
Abstract Text:
This talk presents an efficient methodology for constructing parametric data-driven reduced order models for gyrokinetics. We do so by combining sparse grid interpolation based on (L)-Leja points–to generate the training data and to provide predictions beyond training via interpolation–with optimized dynamic mode decomposition (optDMD), to construct reduced models for each input parameter instance. Outer-loop applications such as stellarator optimization, uncertainty quantification, and real-time control require rapid predictions of high-fidelity, first principles models across large input parameter spaces. Parametric data-driven reduced-order models (ROMs) can provide such evaluations, but standard training data generation is computationally prohibitive due to the curse of dimensionality, with costs scaling exponentially in the number of inputs. The (L)-Leja points are nested with slow growth, yielding sparse grids of low cardinality in low-to-medium dimensional settings–well suited to problems where each training sample requires an expensive forward simulation. The resulting ROMs predict the full 5D gyrokinetic distribution function, from which physically meaningful quantities such as instability growth rates, frequencies, and electromagnetic field fluctuations can be cheaply computed at parameter instances beyond training. We demonstrate the approach in two gyrokinetic micro-instability settings using the GENE code. First, the Cyclone Base Case benchmark demonstrates optDMD ROM prediction capabilities beyond training time horizons and across variations in the binormal wave number. Second, for a six-parameter electron temperature gradient driven micro-instability scenario, we show that a parametric optDMD ROM that generalizes beyond training can be constructed from only 28 high-fidelity simulations via sparse grids. These results highlight the potential of sparse grid-based parametric ROMs to enable otherwise intractable many-query gyrokinetic analyses.
Characterization: 6.0
Comments: