UeFormer: A Unifying Encoding Transformer Neural Operator for meteorological downscaling

Meteorological change presents more and more significant challenges to accurate meteorological forecasting. Traditional meteorological forecasting models, constrained by their fixed-resolution capabilities, often fail to provide precise information for specific coordinates meteorological state. To overcome this limitation, downscaling tasks have gained prominence in meteorology. Operator learning is particularly effective for downscaling because it can capture the partial differential equation (PDE) characteristics of meteorological processes, enabling predictions of meteorological states at arbitrary coordinates. However, conventional operator learning methods often neglect the integration of multimodal meteorological data and do not effectively encode the relationship between query coordinates and the current meteorological state, leading to suboptimal predictions. To address these issues, we propose a novel operator learning method, the Unifying encoding Transformer Neural Operator (Ueformer). This method harnesses the powerful multimodal fusion capabilities of transformers and jointly encodes both query coordinates and meteorological states. Our approach demonstrates superior performance across various meteorological variables and metrics, outperforming baseline methods. Furthermore, we provide theoretical proofs of the method’s approximation capabilities, contributing significantly to the field of meteorology.