A 3D spatial self-attention module on a non-uniform vertical coordinate for super-resolution wind fields

Abstract

Winds are the major factor in the transport of methane (CH4) particles and the accurate advection and turbulence computation in weather simulations is an important task in building chemical transport digital twins. However, the complexity of computation in resolving high-resolution wind fields left the simulations impractical under different conditions and times. Furthermore, the high-resolution weather simulations calculate wind components on a non-uniform vertical coordinate for computational stability, introducing an additional complexity for constructing modeling frameworks using common deep neural networks (DNNs). This study presents a preliminary work to super-resolve the coarse three-dimensional (3D) wind fields, a horizontal 900 m resolution of data, on a non-uniform vertical coordinate from the Weather Research & Forecasting Model (WRF) into the 100 m horizontal resolution of 3D wind fields. For the super-resolution of wind velocity data, we introduce an adversarial DNN that combines convolutional neural networks (CNNs) and a novel spatial self-attention computation module. Our computation module learns horizontal wind velocity structure with 2D convolutional filters but learns the association of wind velocity with adjacent vertical layers based on self-attention computation instead of applying 3D convolution filters to capture better 3D wind information from data. We also introduce a sparse regularization to a self-attention map to amplify and suppress elements in the self-attention map based on the relationship between vertical layers during training. We then train the model with one month of WRF simulation on a single CH4 emission site in the United States to evaluate our approach. The preliminary results demonstrate that our super-resolution model reduces the reconstruction errors increasing along with vertical layers due to the non-uniform vertical grid, and reproduces more realistic wind velocity fields than standard CNNs.