[Paper Review] Physics-Driven Convolutional Autoencoder Approach for CFD Data Compressions

Alberto Olmo, Ahmed S. Zamzam+2 more|arXiv (Cornell University)|Oct 17, 2022

Lattice Boltzmann Simulation Studies4 citations

TL;DR

This paper proposes a physics-driven convolutional autoencoder that improves CFD data compression by incorporating incompressibility and enstrophy preservation into the loss function. The method achieves up to 12x faster training, better physical fidelity, and enhanced explainability without sacrificing reconstruction quality.

ABSTRACT

With the growing size and complexity of turbulent flow models, data compression approaches are of the utmost importance to analyze, visualize, or restart the simulations. Recently, in-situ autoencoder-based compression approaches have been proposed and shown to be effective at producing reduced representations of turbulent flow data. However, these approaches focus solely on training the model using point-wise sample reconstruction losses that do not take advantage of the physical properties of turbulent flows. In this paper, we show that training autoencoders with additional physics-informed regularizations, e.g., enforcing incompressibility and preserving enstrophy, improves the compression model in three ways: (i) the compressed data better conform to known physics for homogeneous isotropic turbulence without negatively impacting point-wise reconstruction quality, (ii) inspection of the gradients of the trained model uncovers changes to the learned compression mapping that can facilitate the use of explainability techniques, and (iii) as a performance byproduct, training losses are shown to converge up to 12x faster than the baseline model.

Motivation & Objective

• Address the growing challenge of memory and I/O bottlenecks in high-fidelity CFD simulations due to massive data volumes.
• Improve data compression models by embedding physical laws—specifically incompressibility and enstrophy preservation—into the training process.
• Enhance model explainability by leveraging gradient-based visualization techniques on physics-informed autoencoders.
• Reduce training time and data requirements while maintaining or improving reconstruction accuracy compared to baseline autoencoders.

Proposed method

• Design a fully convolutional autoencoder with an encoder (E) and decoder (D) to map full-resolution 3D velocity fields (128×128×128) to a compressed latent space and back.
• Introduce physics-informed regularization terms in the loss function: a divergence-free constraint (λ) to enforce incompressibility and an enstrophy-preserving term (β) to maintain vorticity dynamics.
• Optimize the autoencoder using a combined loss: reconstruction loss (MSE) plus physics-based penalties, with hyperparameters λ and β to balance physical fidelity and accuracy.
• Train the model on 1,300 snapshots of decaying isotropic turbulence with Taylor-scale Reynolds numbers between 65 and 105.
• Utilize 3-channel input (u, v, w velocity components) to allow the network to learn inter-component relationships, reducing model depth and training time.
• Apply Grad-CAM heatmaps on the last convolutional layer to visualize attention patterns and assess model explainability across different training configurations.

Experimental results

Research questions

• RQ1 Can enforcing incompressibility and enstrophy preservation in the autoencoder loss improve physical consistency of compressed CFD data without degrading reconstruction quality?

Key findings

• The physics-informed autoencoder achieves a dissipation ratio of 0.504 at β=10, outperforming the baseline (0.544) and demonstrating superior physical fidelity.
• Training with physics-informed losses reduces training time by up to 12x compared to the baseline, even when accounting for 3-channel input processing.
• The model reaches the same MSE loss (0.045) in half the epochs (50 vs. 100) when trained on 566 samples, indicating faster convergence and data efficiency.
• Gradient-based explainability via Grad-CAM shows that physics-informed models focus on high-magnitude physical regions, while the baseline model exhibits random attention patterns.
• The inclusion of physics terms improves enstrophy and dissipation ratio in 6 out of 8 models, confirming consistent physical behavior across hyperparameter settings.
• The model maintains or improves reconstruction quality (MSE) while significantly enhancing physical consistency and training efficiency.