Abstract Details
Abstracts
Author: Sebastian De Pascuale
Requested Type: Consider for Invited
Submitted: 2022-03-04 16:00:40
Co-authors: D. L. Green, J. D. Lore, and K. Allen
Contact Info:
Oak Ridge National Laboratory
1 Bethel Valley Rd
Oak Ridge, Tennessee 37830
United States
Abstract Text:
In this talk, we present a data-driven approach for extracting steady-state dynamics latent around an operating point of SOLPS-ITER simulations. We adapt the projective integration framework to accelerate the prediction of key variables in the tokamak plasma boundary. Linear time advance operators constructed via dynamic mode decomposition (DMD) are used to project profiles of the divertor target density and temperature with stable large time steps and robust error convergence. We show that these operators can be analyzed in terms of spectral components that allow for the identification of advantageous time step stability constraints and the separation of fast and slow timescales for integration. By selectively choosing subset time intervals in the ongoing SOLPS-ITER simulation for stable projection, we demonstrate that the error incurred by this approach increases linearly with speedup factor. We show that these operators can achieve up to an 8x acceleration of the target quantities with an average relative error on the order of 10%. For the SOLPS-ITER simulations considered here, these results are found to be robust to the level of numerical noise. We also address one of the limitations of DMD due to the poor scaling of the SVD on systems with higher degrees of freedom by investigating an alternative low-rank approximation scheme. The algorithm we have applied to SOLPS-ITER data is shown to be at least two orders of magnitude faster than the SVD and capable of data compression down to 1% with a total relative error on the order of 10^-4. These findings encourage further development of data-driven projective integration to overcome the weeks to months wall-clock time of SOLPS-ITER at full fidelity and to extend this work towards more challenging multiscale problems such as those captured by gyrokinetic simulation.
Comments: