Source-linked AI summary
CFDNet: a deep learning-based accelerator for fluid simulations
Octavi Obiols-Sales, Abhinav Vishnu, Nicholas Malaya, Aparna Chandramowlishwaran
TL;DR
CFD simulations are computationally expensive for design optimization, motivating faster methods that preserve high-fidelity convergence constraints. CFDNet couples a CNN with a RANS physics solver to predict key flow variables and refine them physically, achieving 1.9−7.4× speedups while meeting convergence constraints across laminar and turbulent flows, including unseen geometries.
Problem
CFD simulations are computationally expensive, while existing deep-learning approaches can lack conservation-law satisfaction, turbulent-flow support, and generalization to unseen geometries.
Method
CFDNet couples a convolutional neural network predicting velocity, pressure, and eddy viscosity with iterative refinement by the RANS physics solver.
Results
1.9−7.4× speedups were achieved on steady laminar and turbulent flows while preserving the physics solver’s convergence constraints across varied geometries, including unseen ones.
Takeaways & Limitations
Coupling physical models with data-driven models can accelerate simulation convergence while retaining the convergence constraints of the domain-specific solver.
Takeaways & Limitations
The evaluation focuses on fluid-flow simulations, particularly RANS models, and the method remains to be widely used for predictive engineering.
Abstract
from arXiv · showhide
CFD is widely used in physical system design and optimization, where it is used to predict engineering quantities of interest, such as the lift on a plane wing or the drag on a motor vehicle. However, many systems of interest are prohibitively expensive for design optimization, due to the expense of evaluating CFD simulations. To render the computation tractable, reduced-order or surrogate models are used to accelerate simulations while respecting the convergence constraints provided by the higher-fidelity solution. This paper introduces CFDNet -- a physical simulation and deep learning coupled framework, for accelerating the convergence of Reynolds Averaged Navier-Stokes simulations. CFDNet is designed to predict the primary physical properties of the fluid including velocity, pressure, and eddy viscosity using a single convolutional neural network at its core. We evaluate CFDNet on a variety of use-cases, both extrapolative and interpolative, where test geometries are observed/not-observed during training. Our results show that CFDNet meets the convergence constraints of the domain-specific physics solver while outperforming it by 1.9 - 7.4x on both steady laminar and turbulent flows. Moreover, we demonstrate the generalization capacity of CFDNet by testing its prediction on new geometries unseen during training. In this case, the approach meets the CFD convergence criterion while still providing significant speedups over traditional domain-only models.
1 INTRODUCTION
CFDNet couples deep learning with a physics solver to accelerate Reynolds Averaged Navier-Stokes simulations while preserving convergence constraints. It targets limitations in existing approaches, including incomplete conservation, turbulence support, and generalization across geometries.
- CFD simulations solve discretized Navier-Stokes equations iteratively, but their computational expense limits design optimization and motivates acceleration efforts.
- Existing deep-learning approaches may fail to satisfy conservation laws, support turbulent flows, or generalize beyond training geometries.The surveyed state-of-the-art methods satisfy up to six of nine evaluated features.
- CFDNet combines domain-specific knowledge with deep learning to accelerate RANS convergence while meeting the physical solver’s convergence constraints.The framework couples a learned model with the original physics solver rather than replacing the solver outright.
- The framework uses a multi-channel representation and a six-layer encoder-decoder CNN with domain-specific activations to predict velocity, pressure, and eddy viscosity.Training uses intermediate RANS iterations as inputs and domain-specific solver outputs as targets.
- 1.9−7.4× speedup is reported for laminar and turbulent flows across multiple geometries, including cases with geometries unseen during training.The evaluation includes same-geometry, interpolative subset-geometry, and challenging unseen-geometry scenarios.
2 BACKGROUND
The background presents steady incompressible RANS as an approximate time-averaged model for turbulent flow and explains how SIMPLE iteratively solves its discretized equations to convergence.
- Steady incompressible RANS provides an approximate time-averaged solution to the incompressible Navier-Stokes equations for turbulent flows.
- The model uses mean velocity, kinematic mean pressure, fluid viscosity, and eddy viscosity as core flow quantities.The Spalart-Allmaras model supplies a transport equation for modified eddy viscosity.
- Equations (1), (2), and (3) form four PDEs in 2D and five PDEs in 3D, solved numerically on a structured grid with boundary conditions.Spatial derivatives are computed using finite-difference methods and interpolation at grid-cell faces.
- SIMPLE iteratively computes velocity, pressure, and eddy-viscosity fields across the grid until the convergence criterion is satisfied.The algorithm updates intermediate pressure and velocity fields, computes a pressure correction, and checks the residual.
- The physics solver is the iterative procedure used to reduce the number of RANS simulation iterations and obtain the steady-state solution.
3 CFDNET: A DL ACCELERATOR FOR FLUID SIMULATIONS
CFDNet learns a map from intermediate RANS flow fields to converged solutions, using a CNN inside a physics-solver pipeline. Warmup and iterative refinement incorporate domain knowledge and enforce convergence and conservation constraints.
- Model formulation: CFDNet learns a function G that maps an intermediate solver state x_I to the converged state x_N.The state tensors contain z flow variables over an m × n grid.
- Input/output representation: The model represents velocity, pressure, and modified eddy viscosity as spatially arranged channels in a shared CNN input and output tensor.For 2D incompressible RANS, the four variables are the two velocity components, pressure, and modified eddy viscosity.
- Input/output representation: Non-dimensionalization rescales variables with different physical ranges and reduces the number of free parameters for learning.The procedure uses configuration-specific reference values for velocity, pressure, and eddy viscosity.
- Network design and architecture: The six-layer CNN uses three convolution layers followed by three deconvolution layers to predict all steady-state flow variables.The convolution-deconvolution structure is selected to extract spatial correlations within and among fluid-variable channels.
- CFDNet framework: CFDNet couples physics-solver warmup, CNN inference, and iterative refinement rather than operating as a pure surrogate.The refinement stage feeds the CNN output back to the physics solver to satisfy convergence constraints and conservation laws.
- Convergence criteria and error: The framework addresses surrogate limitations by requiring residual convergence and conservation properties, while recognizing that CNN optimization alone does not enforce conservation laws.The paper notes that ground-truth-based error evaluation can be inapplicable when ground-truth solutions are unavailable for unseen cases.
4 EXPERIMENT SETUP
CFDNet is evaluated on wall-bounded and external-aerodynamic flows, using varied Reynolds numbers and geometries to test interpolation, extrapolation, and generalization. The experiments train CNN models from intermediate RANS states paired with converged solutions, across turbulent and laminar settings.
- Case studies: Two flow classes stress different aspects: channel flows vary configurations at fixed geometry, while external aerodynamics vary geometry at fixed flow conditions.The study includes turbulent channel flow and flow around solid bodies under turbulent and laminar conditions.
- Wall-bounded flows: The channel-flow design space is scoped to 4000 < Re < 15000, with training at Re = 4200, 6800, 7500, and 12500.Re = 5600 and Re = 13750 are selected for interpolation and extrapolation tests, respectively.
- External aerodynamics: External-aerodynamics training uses six ellipses spanning aspect ratios 0.1–0.7 at Re = 6e5 for turbulent dataset B and Re = 30 for laminar dataset C.Testing includes an unseen ellipse with AR = 0.3, an airfoil, and a cylinder.
- Dataset creation and preprocessing: The dataset recipe snapshots every intermediate RANS iteration and maps each pre-convergence state independently to the final steady-state solution.Initial conditions are also added as samples, and all samples are non-dimensionalized using flow-configuration-specific reference values.
- Training: The training sets contain 6372, 14953, and 12988 samples for datasets A, B, and C, respectively, with three independently trained CNN models.Training uses Keras with TensorFlow 1.13, RMSProp, mean squared error, and a learning rate of 7e−5.
- Training: Different geometries create distinct flow regimes, including recirculation around thicker ellipses, increasing the learning difficulty relative to the fixed-geometry channel case.The thicker ellipses have more accentuated geometry gradients and more complex nonlinear flow behavior.
5 RESULTS AND DISCUSSION
CFDNet combines CNN predictions with warmup and iterative physics-based refinement to satisfy solver convergence constraints. It accelerates simulations across observed and unseen geometries, although standalone CNN accuracy is weaker for unseen geometries.
- Use cases: CFDNet evaluates observed-geometry and unseen-geometry cases across interpolative and extrapolative flow settings.
- Accuracy: The framework compares CNN predictions and refined CFDNet outputs against OpenFOAM using flow-field errors and conservation-of-mass checks.
- Accuracy: RME is below 2% for all flow variables in the observed-geometry channel-flow case, where CNN prediction errors are 2–3 orders of magnitude below variable values.
- Accuracy: Iterative refinement is necessary because the standalone CNN output is not divergence-free, whereas final CFDNet fields satisfy conservation laws and solver convergence constraints.
- Performance: 1.9–7.4× speedups are achieved across test cases, with CFDNet requiring fewer iterations than the physics solver even for generalized geometries.
- Performance: 1.9–2.6× speedups remain for the DG-SF cylinder despite nonlinear rear-flow behavior and greater refinement requirements.
- Warmup: Without warmup, speedups fall to 1.2–2.3×, and the turbulent-cylinder case can become slower than the physics solver.
- Warmup: Warmup reduces overall refinement requirements and contributes to 1.9–4.6× speedups, while consuming only 1–2% of physics-solver iterations.
6 RELATED WORK
Prior work uses surrogates and neural networks to accelerate fluid simulations, but important limitations remain in generalizability, turbulence coverage, and complete physical-field prediction.
- Gaussian-process and regression surrogates can reduce optimization cost but depend on geometry parameterizations or predict fields only at selected surface points.
- Neural-network approaches accelerate Eulerian, laminar, or viscous-flow simulations through solver replacement or learned mappings.
- CFDNet is positioned against these limitations by targeting broader fluid-field prediction and accelerated simulation across flow settings.
- Laminar-flow methods remain difficult to extend to turbulent flows because turbulence is intrinsically chaotic and harder to resolve.
- Recent turbulent-flow models often predict only subsets of the flow state, such as eddy viscosity without velocity and pressure or without eddy viscosity.
7 CONCLUSIONS
CFDNet couples RANS fluid-flow simulation with a convolutional neural network to accelerate convergence while preserving the physics solver’s convergence constraints. The paper also reports accurate predictions on unseen geometries, while identifying uncertainty quantification and generalizability as open issues.
- 1.9–7.4×: CFDNet accelerates convergence on steady laminar and turbulent flows across varied geometries without relaxing the physics solver’s convergence constraints.
- CFDNet demonstrates accurate predictions on geometries unseen during training, supporting its capacity for generalization and extrapolation.
- The approach is presented as promising for coupling physical models with data-driven machine learning to accelerate simulation convergence.
- Generalizability of the network architecture remains an important open question beyond the paper’s focus on fluid-flow simulations and RANS models.
- Further research is needed to quantify uncertainty and robustness because users may not run a traditional simulation in parallel to verify CFDNet results.