Sparse Convolution-based Markov Models for Nonlinear Fluid Flows

TL;DR

This paper investigates sparse convolution-based Markov models for nonlinear fluid flows, focusing on data-driven feature space mappings that simplify dynamics and enable accurate, real-time predictions.

Contribution

It introduces and compares two novel sparse convolution operators—POD-convolution and sparse Gaussian Process convolution—for improved nonlinear fluid flow modeling.

Findings

POD-convolution effectively captures dominant flow features.

Sparse Gaussian Process convolution offers flexible, implicit feature mappings.

Both methods enhance real-time prediction accuracy in fluid dynamics.

Abstract

Data-driven modeling for nonlinear fluid flows using sparse convolution-based mapping into a feature space where the dynamics are Markov linear is explored in this article. The underlying principle of low-order models for fluid systems is identifying convolutions to a feature space where the system evolution (a) is simpler and efficient to model and (b) the predictions can be reconstructed accurately through deconvolution. Such methods are useful when real-time models from sensor data are needed for online decision making. The Markov linear approximation is popular as it allows us to leverage the vast linear systems machinery. Examples include the Koopman operator approximation techniques and evolutionary kernel methods in machine learning. The success of these models in approximating nonlinear dynamical systems is tied to the effectiveness of the convolution map in accomplishing both (a) and (b) mentioned above. To assess this, we perform in-depth study of two classes of sparse convolution operators: (i) a pure data-driven POD-convolution that uses left singular vectors of the data snapshots - a staple of Koopman approximation methods and (ii) a sparse Gaussian Process (sGP) convolution that combines sparse sampling with a Gaussian kernel embedding an implicit feature map to an inner product reproducing kernel Hilbert space.