Enhancing Computational Fluid Dynamics with Machine Learning

Computational fluid dynamics (CFD) has long served as a vital tool for scientific and industrial flow analysis. However, resolving the chaotic, multi-scale nature of turbulent flows governed by the Navier-Stokes equations remains computationally demanding and, at high Reynolds numbers, often intractable. This paper explores how machine learning (ML) is being integrated into scientific computing to overcome these traditional limitations.

The most computationally expensive aspects of CFD include direct numerical simulation (DNS) and solving the Poisson equation for pressure-velocity coupling. Recent advances leverage neural networks to learn better spatial discretization schemes or to act as correction operators that bridge the gap between coarse-grid simulations and high-fidelity results. By utilizing super-resolution techniques or deep-learning-based pressure solvers, researchers can significantly reduce the computational cost of achieving high-fidelity flow details without requiring exhaustive mesh refinement.

Because resolving all scales is often impossible, industrial applications rely on Reynolds-averaged Navier-Stokes (RANS) or large-eddy simulation (LES) models, which introduce empirical assumptions to represent unresolved turbulent effects. ML is being used to develop more accurate closure models for RANS and subgrid-scale (SGS) models for LES. By training networks on high-fidelity data—or using reinforcement learning agents that adaptively adjust model coefficients—these data-driven approaches aim to handle complex geometries and pressure gradients better than traditional algebraic models.

Reduced-order models (ROMs) provide fast, surrogate representations of complex fluid systems, essential for optimization and control tasks. Traditional linear methods like Proper Orthogonal Decomposition (POD) are being superseded by deep nonlinear autoencoders that learn latent coordinate systems on curved manifolds, offering superior compression. Combined with regression techniques like sparse identification of nonlinear dynamics (SINDy), these models can effectively approximate the evolution of coherent flow structures while maintaining physical interpretability.