Scalable resolvent analysis for three-dimensional flows

TL;DR

This paper introduces RSVD-Δt, a scalable algorithm combining randomized SVD and optimized time-stepping, enabling efficient resolvent analysis of large 3D turbulent flows with reduced computational costs.

Contribution

The paper presents RSVD-Δt, a novel, scalable method that significantly reduces CPU and memory requirements for resolvent analysis in large three-dimensional flow systems.

Findings

RSVD-Δt scales linearly with problem size.

Validated on Ginzburg-Landau and turbulent jet flows.

Enabled analysis of complex flow phenomena previously computationally infeasible.

Abstract

Resolvent analysis is a powerful tool for studying coherent structures in turbulent flows. However, its application beyond canonical flows with symmetries that can be used to simplify the problem to inherently three-dimensional flows and other large systems has been hindered by the computational cost of computing resolvent modes. In particular, the CPU and memory requirements of state-of-the-art algorithms scale poorly with the problem dimension, \ie the number of discrete degrees of freedom. In this paper, we present RSVD-$\Delta t$, a novel approach that overcomes these limitations by combining randomized singular value decomposition with an optimized time-stepping method for computing the action of the resolvent operator. Critically, the CPU cost and memory requirements of the algorithm scale linearly with the problem dimension. We develop additional strategies to minimize these costs and control errors. We validate the algorithm using a Ginzburg-Landau test problem and demonstrate RSVD-$\Delta t$'s low cost and improved scaling using a three-dimensional discretization of a turbulent jet. Lastly, we use it to study the impact of low-speed streaks on the development of Kelvin-Helmholtz wavepackets in the jet via secondary stability analysis, a problem that would have been intractable using previous algorithms.