[Paper Review] DeepCFD: Efficient Steady-State Laminar Flow Approximation with Deep Convolutional Neural Networks

Mateus Dias Ribeiro, Abdul Rehman+2 more|arXiv (Cornell University)|Apr 19, 2020

Model Reduction and Neural NetworksPhysics and Astronomy64 citations

TL;DR

DeepCFD uses a U-Net CNN to approximate 2D non-uniform steady-state laminar flow solutions (velocity and pressure) from boundary geometry, achieving up to 1,000x CPU speedup over conventional CFD with low error. It extends prior work to provide complete velocity and pressure fields with multiple decoders.

ABSTRACT

Computational Fluid Dynamics (CFD) simulation by the numerical solution of the Navier-Stokes equations is an essential tool in a wide range of applications from engineering design to climate modeling. However, the computational cost and memory demand required by CFD codes may become very high for flows of practical interest, such as in aerodynamic shape optimization. This expense is associated with the complexity of the fluid flow governing equations, which include non-linear partial derivative terms that are of difficult solution, leading to long computational times and limiting the number of hypotheses that can be tested during the process of iterative design. Therefore, we propose DeepCFD: a convolutional neural network (CNN) based model that efficiently approximates solutions for the problem of non-uniform steady laminar flows. The proposed model is able to learn complete solutions of the Navier-Stokes equations, for both velocity and pressure fields, directly from ground-truth data generated using a state-of-the-art CFD code. Using DeepCFD, we found a speedup of up to 3 orders of magnitude compared to the standard CFD approach at a cost of low error rates.

Motivation & Objective

• Motivate reducing CFD computation cost for steady laminar flows in engineering design and optimization.
• Propose a CNN surrogate (DeepCFD) that learns velocity and pressure fields from CFD-ground-truth data.
• Extend previous approaches to output complete velocity and pressure fields using a U-Net architecture with separate decoders.
• Provide dataset and code to enable reproducibility and community contribution.

Proposed method

• Represent flow geometry using a signed distance function (SDF) and a multiple-class flow region channel as CNN inputs.
• Use a down-sampling encoder to learn a latent geometry representation (LGR), followed by up-sampling decoders to map LGR to velocity (Ux, Uy) and pressure (p).
• Compare Autoencoder (AE) baselines with U-Net based architectures and multiple decoders (AE-3, UNet-3, etc.).
• Train with ground-truth CFD data generated by OpenFOAM (simpleFoam) for 2D channel flow around randomly shaped obstacles.
• Use AdamW optimizer with batch size 64; loss combines mean squared error for velocities and mean absolute error for pressure; normalize losses to avoid bias.
• Hyper-parameter search across learning rates, kernel sizes, encoder/decoder depths, and normalization choices.

Experimental results

Research questions

• RQ1 Can a CNN surrogate accurately approximate 2D non-uniform steady-state laminar flow fields (Ux, Uy, p) from geometric boundary conditions?
• RQ2 Does a U-Net with separate decoders provide better velocity and pressure reconstruction than a baseline autoencoder with shared decoders?
• RQ3 What is the accuracy and speedup trade-off of DeepCFD compared to conventional CFD (OpenFOAM) and a baseline model across diverse obstacle geometries?
• RQ4 How does DeepCFD perform in terms of velocity/pressure field reconstruction versus ground-truth CFD across qualitative flow features (stagnation, separation, pressure zones)?

Key findings

• DeepCFD (UNet-3 architecture) achieves the best test MSE across Ux, Uy, and p compared with the baseline and other architectures.
• For 5-sample tests, UNet-1 vs DeepCFD shows DeepCFD reducing total MSE to 2.0303 ± 0.1360 compared with UNet-1’s 2.9203 ± 0.1520.
• Compared to the baseline AE-1, DeepCFD improves total MSE from 4.4981 ± 0.1753 to 2.0303 ± 0.1360 on the test set, with corresponding reductions in Ux, Uy, and p errors.
• Relative to ground-truth CFD, DeepCFD maintains low error while delivering up to about 3 orders of magnitude speedup on CPU-CPU and up to 5 orders of magnitude with GPU-CPU for batch predictions.
• Qualitative flow reconstructions show DeepCFD capturing key features such as high-pressure frontal regions and flow separation for various obstacle shapes.
• Prediction error distributions indicate most predictions lie at low relative error (often under 10%).
• DeepCFD provides complete velocity and pressure fields, not just velocity magnitude, enhancing utility for engineering applications.