Data-assisted, physics-informed propagators for recurrent flows

Abstract

A novel approach to simulate dynamic, recurrent flows with very large time steps is presented. Data-driven forecasts based on the method of analogues are made employing a set of precomputed time series segments of flow fields. These predictions are then corrected with the evolved deviation between initial flow state and its nearest neighbor in the database. The evolution of these deviations is obtained from the convolution with the propagator of the linearized Navier-Stokes equations, where linearization takes place around the most similar state in the database. The error $E_{\text{lin}}$ due to this approximation implies the appropriate distance function to determine nearest-neighbor fields. We tested the method on the simple case of 2D vortex shedding behind a cylinder at Reynolds numbers in the range of 200 with step sizes of up to 200 times larger than those used to generate the underlying data with the PISO algorithm. Short-term forecasts showed a high degree of accuracy with errors correlating well with $E_{\text{lin}}$ for a number of case variations. But also after many steps, the predicted flow fields remained in close agreement with high-fidelity CFD calculations with relative errors on the order of 1% for the temporal averages. Furthermore, excellent conservation behavior of global properties like total momentum and kinetic energy was observed.

References in corpus (18)

• Discovering governing equations from data: Sparse identification of nonlinear dynamical systems
• DeepONet: Learning nonlinear operators for identifying differential equations based on the universal approximation theorem of operators
• Machine Learning for Fluid Mechanics
• Hidden Physics Models: Machine Learning of Nonlinear Partial Differential Equations
• Nonlinear mode decomposition with convolutional neural networks for fluid dynamics
• Convolutional neural network based hierarchical autoencoder for nonlinear mode decomposition of fluid field data
• Pod-Galerkin Reduced Order Methods for CFD Using Finite Volume Discretisation: Vortex Shedding Around a Circular Cylinder
• Sparse identification of nonlinear dynamics with low-dimensionalized flow representations
• Data-driven modelling of nonlinear spatio-temporal fluid flows using a deep convolutional generative adversarial network
• Super-resolution reconstruction of turbulent flow at various Reynolds numbers based on generative adversarial networks
• Learning Unknown Physics of non-Newtonian Fluids
• The prediction of future from the past: an old problem from a modern perspective
• Data-driven discovery of Green's functions with human-understandable deep learning
• Sparse nonlinear models of chaotic electroconvection
• Data-driven modeling of the chaotic thermal convection in an annular thermosyphon
• Pressure-robust analysis of divergence-free and conforming FEM for evolutionary incompressible Navier-Stokes flows
• Revealing the state space of turbulence using machine learning
• Data-driven low-dimensional dynamic model of Kolmogorov flow