Bayesian Methods for the Navier-Stokes Equations

We develop a Bayesian methodology for numerical solution of the incompressible Navier–Stokes equations with quantified uncertainty. The central idea is to treat discretized Navier–Stokes dynamics as a state-space model and to view numerical solution as posterior computation: priors encode physical structure and modeling error, and the solver outputs a distribution over states and quantities of interest rather than a single trajectory. In two dimensions, stochastic representations (Feynman–Kac and stochastic characteristics for linear advection–diffusion with prescribed drift) motivate Monte Carlo solvers and provide intuition for uncertainty propagation. In three dimensions, we formulate stochastic Navier–Stokes models and describe particle-based and ensemble-based Bayesian workflows for uncertainty propagation in spectral discretizations. A key computational advantage is that parameter learning can be performed stably via particle learning: marginalization and resample–propagate (one-step smoothing) constructions avoid the weight-collapse that plagues naive sequential importance sampling on static parameters. When partial observations are available, the same machinery supports sequential observational updating as an additional capability. We also discuss non-Gaussian (heavy-tailed) error models based on normal variance-mean mixtures, which yield conditionally Gaussian updates via latent scale augmentation.

💡 Research Summary

This paper introduces a comprehensive Bayesian framework for solving the incompressible Navier‑Stokes equations with quantified uncertainty. The central premise is to reinterpret a discretized Navier‑Stokes solver as a Bayesian state‑space model: the hidden state comprises the velocity (or vorticity) field, the transition kernel encodes the numerical time‑stepping scheme, and priors are placed on initial conditions, physical parameters (viscosity, forcing), and model error. The posterior distribution over the state therefore represents a full probabilistic solution rather than a single deterministic trajectory.

In two dimensions, the vorticity formulation reduces to a linear advection‑diffusion equation. By invoking the Feynman‑Kac theorem, the solution can be expressed as an expectation over stochastic particle paths driven by a prescribed drift (the velocity recovered from vorticity via the Biot‑Savart law) and diffusion √(2ν). This representation motivates a Monte‑Carlo solver: each particle follows the stochastic characteristic, and the ensemble average yields an estimate of the vorticity field. The Bayesian view treats the Monte‑Carlo ensemble as a weighted empirical posterior, allowing natural incorporation of prior information and uncertainty propagation.

In three dimensions, the presence of nonlinearity, non‑local Biot‑Savart coupling, and vortex‑stretching precludes a closed‑form Feynman‑Kac formula. The authors therefore adopt the Constantin‑Iyer stochastic Lagrangian representation, where a random flow map X(α,t) satisfies dX = u(X,t)dt + √(2ν)dW_t and the velocity is recovered as an expectation involving the inverse map A_t = X_t^{-1}. This yields an implicit fixed‑point problem that can be tackled with interacting‑particle or ensemble methods. The paper details particle‑based and ensemble‑Kalman‑type algorithms designed for high‑dimensional spectral discretizations, emphasizing techniques such as Rao‑Blackwellization, marginalization of static parameters, and low‑dimensional latent augmentations to mitigate weight degeneracy.

A major contribution is the development of particle learning for online parameter estimation. Standard sequential importance sampling suffers from weight collapse when static parameters (e.g., viscosity, forcing coefficients) are included. By maintaining sufficient statistics and performing a resample‑propagate (one‑step smoothing) step, the algorithm marginalizes over parameters and updates them directly from the particle set, achieving stable learning without collapse. This approach seamlessly extends to sequential data assimilation: when partial observations of the flow are available, the same particle/ensemble machinery incorporates a likelihood term and produces a filtering distribution for the hidden state.

To address the heavy‑tailed error structures typical of turbulent flows, the authors propose normal variance‑mean mixture models (Student‑t, Laplace, Normal‑Inverse‑Gaussian). Introducing a latent scale variable renders the conditional updates Gaussian, preserving computational tractability while providing robustness against outliers.

The paper concludes with illustrative applications in geophysical fluid dynamics (e.g., oceanic and atmospheric modeling) and flow control, where uncertainty quantification is essential. It highlights current challenges—such as the computational burden of high‑dimensional particle systems and the need for scalable resampling—and suggests future directions, including variational Bayesian approximations and deep‑learning‑informed priors. Overall, the work establishes a unified Bayesian perspective that transforms Navier‑Stokes solvers into probabilistic inference engines capable of delivering uncertainty‑aware predictions and adaptive learning.