Data-driven global weather predictions at high resolutions

Abstract

Society has benefited enormously from the continuous advancement in numerical weather prediction that has occurred over many decades driven by a combination of outstanding scientific, computational and technological breakthroughs. Here, we demonstrate that data-driven methods are now positioned to contribute to the next wave of major advances in atmospheric science. We show that data-driven models can predict important meteorological quantities of interest to society such as global high resolution precipitation fields (0.25°) and can deliver accurate forecasts of the future state of the atmosphere without prior knowledge of the laws of physics and chemistry. We also show how these data-driven methods can be scaled to run on supercomputers with up to 1024 modern graphics processing units and beyond resulting in rapid training of data-driven models, thus supporting a cycle of rapid research and innovation. Taken together, these two results illustrate the significant potential of data-driven methods to advance atmospheric science and operational weather forecasting.

A Deep Learning Model for Forecasting Global Monthly Mean Sea Surface Temperature Anomalies

“…Given the results presented here and the successful application of the Unet-LSTM model in previous studies (Taylor, Larraondo, and de Supinski 2021) we have increased confidence that the Unet-LSTM model can be applied to the general problem of the spatial and temporal evolution of other 2D geophysical fields. As of TensorFlow 2.6, a ConvLSTM3D layer is now available.…”

“…In order to efficiently load the model training data using a data-parallel approach we distribute the model data required by each GPU onto the CPU memory of the corresponding node, as we have done in prior studies (Taylor, Larraondo, and de Supinski 2021). The data required on each GPU is read once from a single NetCDF file containing the preprocessed data as described above.…”

“…Previous work by (Larraondo, Renzullo, Inza, and Lozano 2019) investigated the application of SegNet, VGG16 and U-net to the prediction of precipitation fields and concluded that U-net delivered the best estimates of precipitation while employing significantly fewer model parameters. Based on these advantages and our successful application of our U-net based model in previous studies (Taylor, Larraondo, and de Supinski 2021)to the prediction of surface precipitation and forecasting 500 hPa geopotenial height, we have adopted the Unet-LSTM model (Taylor, Larraondo, and de Supinski 2021), as the underlying model architecture for our study. The Unet-LSTM model code is available here (Taylor 2021).…”

“…In this study, we propose a new deep learning modeling framework to forecast monthly global SST, using a Unet-LSTM convolutional encoder-decoder neural network (Taylor, Larraondo, and de Supinski 2021), which has been proven to have better prediction skills while using fewer parameters, compared with other deep learning architectures (Larraondo, Renzullo, Inza, and Lozano 2019). We train the model with observed (reanalysis) SST and surface air temperature data over the past 7 decades, to demonstrate potential long lead predictions for SST variability in the tropical-subtropical oceans.…”

A deep learning model for forecasting global monthly mean sea surface temperature anomalies

“…In this study, we propose a new deep learning modeling framework to forecast monthly global SST, using an Unet-LSTM convolutional encoder-decoder neural network (Taylor et al, 2021), which has been proven to have better prediction skills while using fewer parameters, compared with other deep learning architectures (Larraondo et al, 2019). We train the model with observed (reanalysis) SST and surface air temperature data over the past 7 decades to demonstrate potential long lead predictions for SST variability in the tropicalsubtropical oceans.…”