Replacing Numerical Weather Prediction with Deep Learning and Extenstions to a Full Earth-System Model

We compare the performance of a global deep-learning weather prediction (DLWP) model with reanalysis data and forecasts from the European Center for Medium Range Weather Forecasts (ECMWF). The model is trained using ECMWF ReAnalysis 5 (ERA5) data with convolutional neural networks (CNNs) on a HEALPix mesh using a loss function that minimizes forecast error over a single 24-hour period. The model predicts seven 2D shells of atmospheric data on an equal-area pixelization at resolutions of roughly 200 km. Notebly, our model can bee iterated forward indefinitely to produce forecasts at 6-hour temporal resolution for any lead time. We present case studies showing the extent to which the model is able to reproduce the dynamical evolution of atmospheric systems and its ability to learn "model physics" to forecast two-meter temperature and precipitation. Sources of ensemble spread and the performance of the ensemble are discussed relative to the ECMWF S2S ensemble forecasts. Extensions to full earth-system model are presented using similar deep learning architecture to forecast sea surface temperatures. The SST model can be stably stepped forward for a year and shows skills in forecasting El Ninos. 

About this series: The PAOC Colloquium is a weekly interdisciplinary seminar series that brings together the whole PAOC community. Seminar topics include all research concerning the physics, chemistry, and biology of the atmospheres, oceans and climate, but also talks about e.g. societal impacts of climatic processes. The seminars take place on Monday from 12-1pm. Contact paoc-colloquium-comm@mit.edu for more information and Zoom password.