Abstract
Physics-informed neural networks (NN) are an emerging technique to improve spatial resolution and enforce physical consistency of data from physics models or satellite observations. A super-resolution (SR) technique is explored to reconstruct high-resolution images ($4\times$) from lower resolution images in an advection-diffusion model of atmospheric pollution plumes. SR performance is generally increased when the advection-diffusion equation constrains the NN in addition to conventional pixel-based constraints. The ability of SR techniques to also reconstruct missing data is investigated by randomly removing image pixels from the simulations and allowing the system to learn the content of missing data. Improvements in S/N of $11\%$ are demonstrated when physics equations are included in SR with $40\%$ pixel loss. Physics-informed NNs accurately reconstruct corrupted images and generate better results compared to the standard SR approaches.
Introduction
High-fidelity physical simulations, such as those tracking atmospheric pollutant transport, are computationally intensive. To mitigate this, researchers often utilize coarse grids, which unfortunately results in a loss of critical spatial detail. While standard Super Resolution (SR) techniques from computer vision can upscale these images, they frequently produce results that, while visually plausible, are physically inconsistent—for instance, violating mass conservation or ignoring wind vectors.
Furthermore, real-world observational data, such as satellite imagery, is often incomplete due to cloud cover or sensor malfunctions. Traditional SR methods struggle to fill these gaps while maintaining the underlying physical consistency of dynamic phenomena like moving plumes. This paper introduces PINNSR (Physics-Informed Neural Network Super Resolution), a framework designed to reconstruct high-resolution, physically grounded atmospheric maps even from corrupted or incomplete low-resolution data.
Methodology
The PINNSR framework builds upon the Residual-in-Residual Dense Block (RRDB) architecture. Unlike standard SR models that often use Generative Adversarial Networks (GANs) to prioritize aesthetic sharpness—sometimes at the cost of physical accuracy—PINNSR focuses on high physical fidelity and signal-to-noise ratio.
The core innovation lies in the dual-loss function:
- Pixel-Based Loss ($L_{pix}$): A standard MSE loss comparing the generated output to the ground truth.
- Physics-Consistency Loss ($L_{phys}$): A novel term based on the advection-diffusion equation. The model calculates the "physics residual"—the degree to which the generated image violates fluid transport laws—and penalizes the network accordingly.
The study utilizes a custom dataset of 2D atmospheric dispersion plumes. Crucially, the low-resolution (LR) inputs are generated by running the simulation on a coarse grid rather than simple downsampling, providing a more realistic and challenging testbed. To handle temporal dynamics, inputs are stacked as three-channel images ($t-1, t, t+1$), enabling the calculation of time derivatives necessary for the physics loss.
Results and Discussion
The inclusion of physics as a regularizer proves highly effective, especially when dealing with missing data. When input pixels are randomly removed (up to 60%), the physics loss guides the network to "fill in the blanks" based on the governing equations of advection and diffusion.
Key findings include:
- Enhanced Accuracy: PINNSR consistently outperforms bicubic interpolation and standard SR methods. At a 40% pixel loss, PINNSR achieved an 11% reduction in RMS error compared to standard SR.
- Physical Reliability: The physics residual of PINNSR's output is significantly lower than that of competing models, indicating that the reconstructed plumes adhere more closely to real-world fluid dynamics.
- Addressing the "Downsampling Trap": The authors demonstrate that models trained solely on synthetic, downsampled data fail to generalize to "native" low-resolution data, whereas PINNSR's physics-based approach remains robust.
Conclusion
PINNSR demonstrates the power of embedding physical laws into deep learning architectures for scientific data reconstruction. While currently limited to 2D models with known parameters (wind velocity and diffusivity), it paves the way for more complex 3D applications and joint parameter-state estimation in Earth sciences. Future work will explore more complex governing equations, such as Navier-Stokes, to handle turbulent atmospheric conditions.
Summary updated