Abstract:Fluid data completion is a research problem with high potential benefit for both experimental and computational fluid dynamics. An effective fluid data completion method reduces the required number of sensors in a fluid dynamics experiment, and allows a coarser and more adaptive mesh for a Computational Fluid Dynamics (CFD) simulation. However, the ill-posed nature of the fluid data completion problem makes it prohibitively difficult to obtain a theoretical solution and presents high numerical uncertainty and instability for a data-driven approach (e.g., a neural network model). To address these challenges, we leverage recent advancements in computer vision, employing the vector quantization technique to map both complete and incomplete fluid data spaces onto discrete-valued lower-dimensional representations via a two-stage learning procedure. We demonstrated the effectiveness of our approach on Kolmogorov flow data (Reynolds number: 1000) occluded by masks of different size and arrangement. Experimental results show that our proposed model consistently outperforms benchmark models under different occlusion settings in terms of point-wise reconstruction accuracy as well as turbulent energy spectrum and vorticity distribution.

What problem does this paper attempt to address?

This paper discusses the use of deep learning to address the problem of completing fluid dynamics data. In experimental and computational fluid dynamics, data completion of fluid is a research problem with high potential benefits as it can reduce the number of sensors required in experiments and allow the use of coarser and adaptive grids in numerical simulations. However, finding a solution to the data completion problem is theoretically challenging due to its inherent nature of undefined problem definition, and data-driven methods such as neural network models can introduce high numerical uncertainty and instability. To address these issues, the paper proposes a two-stage learning method that utilizes vector quantization techniques in computer vision to map complete and incomplete fluid data to discrete low-dimensional representations. This approach first learns the latent representation of the original 2D turbulent data using an autoencoder neural network model, and then fine-tunes the encoding module to predict the unmasked data samples, while employing a fixed-weight decoder to evaluate the prediction error. By using vector quantization, the goal of the paper is to reduce the prediction error caused by quantization. Experimental results show that the proposed model outperforms the baseline models in terms of point accuracy, turbulent energy spectrum, and vortex distribution, under different masking settings. The paper selects 2D Kolmogorov flow data (with a Reynolds number of 1000) as the experimental object, as its distinct turbulent patterns pose a greater challenge to data completion and provide insights into how vortex properties and mask selection affect reconstruction accuracy. Furthermore, the paper compares the model with neural network-based operator learning methods such as Fourier Neural Operator (FNO) and Factorized Transformer (FactFormer). Although neural operators can theoretically be directly used for fluid flow data completion, the experiments found that models using continuous latent space perform worse than those that learn discrete latent space through vector quantization when dealing with undefined data completion problems. In conclusion, this paper proposes an effective method to address the completion of fluid dynamics data by combining deep learning with vector quantization techniques, particularly excelling in handling large-scale continuous missing data regions.

Inpainting Computational Fluid Dynamics with Deep Learning

Computational Fluid Dynamics Iteration Driven by Data

Physics-Based Deep Learning for Flow Problems

HMDST: A Hybrid Model-Data Driven Approach for Spatio-Temporally Consistent Video Inpainting

Recovering missing CFD data for high-order discretizations using deep neural networks and dynamics learning

PiRD: Physics-informed Residual Diffusion for Flow Field Reconstruction

Data recovery in computational fluid dynamics through deep image priors

Hidden Fluid Mechanics: A Navier-Stokes Informed Deep Learning Framework for Assimilating Flow Visualization Data

Physics informed neural fields for smoke reconstruction with sparse data

OIFlow: Occlusion-Inpainting Optical Flow Estimation by Unsupervised Learning

Enhancing High-Resolution Reconstruction of Flow Fields Using Physics-Informed Diffusion Model with Probability Flow Sampling

Feel the Force: From Local Surface Pressure Measurement to Flow Reconstruction in Fluid-Structure Interaction

Shallow Neural Networks for Fluid Flow Reconstruction with Limited Sensors

A practical approach to flow field reconstruction with sparse or incomplete data through physics informed neural network

SlimFliud-Net: Fast Fluid Simulation Using Admm Pruning.

Inferring Fluid Dynamics via Inverse Rendering

Fluid Inverse Volumetric Modeling and Applications from Surface Motion

PINNs4Drops: Convolutional feature-enhanced physics-informed neural networks for reconstructing two-phase flows

Unsupervised Denoising and Super-Resolution of Vascular Flow Data by Physics-informed Machine Learning

Machine learning accelerated computational fluid dynamics

A Fluid Flow Data Set for Machine Learning and its Application to Neural Flow Map Interpolation