Towards a Digital Twin of Atmospheric Flow for Microscale Applications

We require high-resolution information of atmospheric flow on the order of tens of meters or less for a host of applications, including wind energy, urban heat or contaminant dispersion, details of hurricane simulation, modeling wildland and urban interface fires, and more. Although it is becoming common to do increasingly high-resolution simulations on the order of tens of meters or meters) for microscale atmospheric conditions with large-eddy simulation (LES) codes, it remains expensive in terms of computing time, and we are seldom able to perform these simulations faster than real time. Machine learning (ML) downscaling methods, however, are able to mimic such LES runs in a matter of seconds. We report on using Enhanced Super-resolution Generative Adversarial Networks (ESRGANs) to provide a 32-times downscale from mesoscale simulations into the microscale to resolve features in complex terrain, hurricanes, and urban areas. We employ LES data for training, validating, and testing the ML methods and compare the downscaled results in terms of the statistical moments and ability to reproduce the turbulent energy spectrum for the regions of interest. We demonstrate the viability of downscaling from the mesoscale simulations at 1 km resolution to approximately 30 m. For the urban regions, we can downscale from about 160-m to 5-m resolution. We find that physics-inspired cost functions help with the training and that we are able to reconstruct the turbulence parameters. Our vision is to add a neural temporal operator to reproduce the behavior at fine temporal and spatial scales to provide detailed information that can be used for a variety of applications.