This repository contains the source code for the research paper "Physics-Informed Shadowgraph Density Field Reconstruction".

Overview

This code implements a physics-informed framework for reconstructing density fields from shadowgraph images. The approach combines shadowgraph imaging techniques with physics-informed neural networks (PINNs) to capture refractive index variations in complex flow fields accurately.

The video below demonstrates the on-time prediction. Due to the lack of high-performance GPU support on this laptop, the prediction process is relatively slow. (P.S.: Our model was definitely not trained on this laptop! 😊)

Key Features

• Shadowgraph Image Processing: Pre-processing and analysis of shadowgraph images for density field visualization.
• PINN Implementation: Physics-informed neural network setup tailored for accurate density field reconstruction.
• Density Field Reconstruction: Algorithms for computing density distributions based on refractive index variations within the experimental field.

Dataset

The data folder only contains a few example images. Its purpose is to illustrate what the shadowgraph image looks like. Training solely on this data will result in significant overfitting. A complete example dataset can be obtained by contacting the authors.

Requirements

All necessary dependencies are listed in requirements.txt.

Authors

• Primary Author: Xutun Wang, Yuchen Zhang
• Contact Information: You can reach us at [xt-wang24@mails.tsinghua.edu.cn] or through our academic institution profiles.
• Special thanks to Dr. Yuchen Zhang @paradoxknight1 for his significant contributions to this research.

Reference

Please cite this article as:
X. Wang, Y. Zhang, Z. Li, H. Wen, B. Wang, Physics-informed shadowgraph network: an end-to-end self-supervised density field reconstruction method, Experimental Thermal and Fluid Science(2025), doi: https://doi.org/10.1016/j.expthermflusci.2025.111562