A Diffusion Framework for Temperature Fields Reconstruction With Embedded Vertical Dynamics

Abstract

High-resolution meteorological data are crucial for accurate climate research and weather forecasting. However, climate change and global warming are altering the patterns in meteorological data, leading to gradual shifts in the statistical characteristics of climate variables. Existing statistical downscaling approaches primarily focus on fitting historical data while neglecting the integration of essential physical laws governing atmospheric dynamics. Although these models may perform well on past data, they risk losing accuracy when applied to future climate scenarios, particularly when faced with climate shifts or changing conditions. To address this challenge, we introduce a physics-informed framework to downscale continuous temperature fields. Our framework is designed to decompose the temperature fields into a primary deterministic component and stochastic residual component, each modeled by distinct parts of the architecture. The deterministic component reconstructs the primary temperature field, whereas the stochastic diffusion component captures small-scale details and uncertainties. Moreover, this framework integrates vertical dynamics by incorporating physical priors derived from the fundamental temperature variation equation, combined through zero convolution, and applied as physics priors in the downscaling process using a 3D U-Net architecture as the encoder. The model's loss function includes Charbonnier loss for data fitting along with static stability loss, gradient loss, and coupling loss to ensure physical consistency and accurate vertical interaction representation. Comparative experiments demonstrate that our method outperforms traditional techniques, reducing the error between downscaled results and high-resolution observations to 0.627 K compared to 1.394 K with bicubic interpolation.