Chulin Wang,Kyongmin Yeo,Xiao Jin,Andres Codas,Levente J. Klein,Bruce Elmegreen
Abstract:We present a super-resolution model for an advection-diffusion process with limited information. While most of the super-resolution models assume high-resolution (HR) ground-truth data in the training, in many cases such HR dataset is not readily accessible. Here, we show that a Recurrent Convolutional Network trained with physics-based regularizations is able to reconstruct the HR information without having the HR ground-truth data. Moreover, considering the ill-posed nature of a super-resolution problem, we employ the Recurrent Wasserstein Autoencoder to model the uncertainty.
What problem does this paper attempt to address?
The problem that this paper attempts to solve is to perform super - resolution reconstruction and prediction on the advection - diffusion process in the absence of high - resolution (HR) ground - truth data. Specifically:
1. **Lack of high - resolution data**: Most super - resolution models assume the existence of high - resolution ground - truth data during training, but in many practical situations, such high - resolution data are not easily obtainable. This paper proposes a self - supervised method to reconstruct high - resolution information through physical constraints and probability modeling in the absence of high - resolution ground - truth data.
2. **Handling uncertainty**: The super - resolution problem is ill - posed in nature, that is, one low - resolution data may correspond to multiple high - resolution solutions. To handle this uncertainty, the authors introduce the Recurrent Wasserstein Autoencoder (RWAE) to model the uncertainty of the output.
3. **Spatio - temporal super - resolution and prediction**: This method not only achieves super - resolution reconstruction in space but also can predict future time steps, and is suitable for scenarios where both spatial and temporal dimensions need to be considered simultaneously, such as climate/weather modeling, satellite observations, etc.
### Specific problem description
The specific problems mentioned in the paper include:
- **Advection - diffusion equation**: The research object is a two - dimensional advection - diffusion process, and its mathematical expression is:
\[
\frac{\partial c}{\partial t}+\nabla\cdot(c\mathbf{u})=\nabla\cdot(K\cdot\nabla c)+Q
\]
where \( c \) is the concentration, \( \mathbf{u} \) is the velocity field, \( K \) is the eddy diffusion tensor, and \( Q \) is the source term. Since \( Q \) and \( K \) are usually unknown and only low - resolution velocity field and concentration data are available, how to reconstruct high - resolution images from this limited information and make predictions has become a challenge.
- **Dataset generation**: In the experiment, a simulated dataset was used, where the high - resolution grid size was \( N_H = 128 \), and the low - resolution training data was generated by spatial averaging of the high - resolution data, with a grid size of \( N_L = 16 \).
### Method innovation points
- **Self - supervised learning**: It does not require high - resolution ground - truth data and achieves super - resolution through physical constraints and probability modeling.
- **Probability modeling**: The Recurrent Wasserstein Autoencoder (RWAE) is used to model uncertainty, enabling the model to generate multiple possible high - resolution solutions that conform to physical laws.
- **Physical constraints**: By introducing physical regularization terms, it is ensured that the model output conforms to physical laws, such as the mass conservation condition \( \nabla\cdot\mathbf{u} = 0 \).
### Experimental results
The experimental results show that this model outperforms the traditional bicubic interpolation method on multiple evaluation metrics, especially in terms of physical consistency. For example, the physical error \( \epsilon_{\text{div}} \) of the model is one order of magnitude smaller than that of the baseline method, indicating that it has a significant advantage in maintaining physical laws.
In summary, this paper aims to solve how to achieve super - resolution reconstruction and prediction of the advection - diffusion process through self - supervised learning and probability modeling in the absence of high - resolution data and ensure that the model output conforms to physical laws.