[Paper Review] Computing with vortices: Bridging fluid dynamics and its information-processing capability

Ken Goto, Kohei Nakajima+1 more|arXiv (Cornell University)|Jan 23, 2020

Neural Networks and Reservoir Computing38 references4 citations

TL;DR

This paper proposes using the Karman vortex system in fluid dynamics as a physical reservoir computing (PRC) platform, demonstrating that its computational performance—measured by memory capacity and nonlinear function approximation—peaks near the critical Reynolds number at the onset of vortex shedding, challenging the conventional 'edge of chaos' paradigm by linking optimal computation to Hopf bifurcation dynamics.

ABSTRACT

Herein, the Karman vortex system is considered to be a large recurrent neural network, and the computational capability is numerically evaluated by emulating nonlinear dynamical systems and the memory capacity. Therefore, the Reynolds number dependence of the Karman vortex system computational performance is revealed and the optimal computational performance is achieved near the critical Reynolds number at the onset of Karman vortex shedding, which is associated with a Hopf bifurcation. Our finding advances the understanding of the relationship between the physical properties of fluid dynamics and its computational capability as well as provides an alternative to the widely believed viewpoint that the information processing capability becomes optimal at the edge of chaos.

Motivation & Objective

• To investigate the information-processing capability of fluid dynamics, particularly in the context of vortex shedding.
• To evaluate the Karman vortex system as a physical reservoir for reservoir computing (PRC).
• To determine how Reynolds number influences computational performance in fluid-based computation.
• To challenge the widely held belief that optimal computation occurs at the 'edge of chaos' by identifying a physical mechanism—Hopf bifurcation—where performance peaks.
• To demonstrate the feasibility of using Navier-Stokes-based simulations to analyze complex fluid dynamics for machine learning tasks.

Proposed method

• The Karman vortex system, generated by flow past a cylinder, is modeled using the incompressible Navier-Stokes equations with a stabilized finite element method (LG method).
• Input signals are applied via time-varying inflow velocity at the boundary, modulating the vortex dynamics.
• The reservoir is constructed from spatial nodes in the fluid domain, where velocity and pressure fields at each node form the reservoir state.
• Echo State Property (ESP) is evaluated to ensure the system's sensitivity to input and stability of internal dynamics.
• Benchmark tasks—nonlinear function approximation and time-series prediction—are performed using linear readout weights trained on reservoir states.
• Computational performance is quantified via memory capacity and nonlinear function approximation accuracy, analyzed across varying Reynolds numbers.

Experimental results

Research questions

• RQ1 Does the Karman vortex system exhibit sufficient dynamical richness to serve as a reservoir for information processing?
• RQ2 How does the computational performance of the fluid reservoir vary with Reynolds number?
• RQ3 Is there an optimal Reynolds number where computational capability is maximized, and if so, what physical mechanism underlies this peak?
• RQ4 Does the peak performance occur at the 'edge of chaos' or at a different dynamical regime?
• RQ5 Can the Navier-Stokes system be effectively used to simulate and analyze reservoir computing tasks in fluid dynamics?

Key findings

• The Karman vortex system achieves optimal computational performance near the critical Reynolds number at the onset of vortex shedding, which corresponds to a Hopf bifurcation.
• The peak memory capacity and nonlinear function approximation accuracy occur precisely at the Hopf bifurcation point, not at higher Reynolds numbers associated with chaotic dynamics.
• The Echo State Property (ESP) is maintained near the bifurcation point, indicating stable and input-sensitive reservoir dynamics.
• The long-diameter oscillation of the twin vortices is strongly modulated by input signals, providing a measurable dynamical response for information encoding.
• The computational performance degrades at higher Reynolds numbers due to excessive turbulence, which disrupts the predictable and stable reservoir dynamics.
• Numerical simulations using the stabilized LG finite element method successfully capture the detailed spatio-temporal dynamics of vortices and their response to inputs, enabling accurate performance evaluation.