TEEMLEAP—A New Testbed for Exploring Machine Learning in Atmospheric Prediction for Research and Education

Abstract

In the past 5 years, data-driven prediction models and Machine Learning (ML) techniques have revolutionized weather forecasting. Meteorological services around the world are now developing ML components to enhance (or even replace) their numerical weather prediction systems. This shift creates new challenges and opportunities for universities and research centers, calling for a much closer cooperation of meteorology with mathematics and computer sciences, updates of teaching curricula, and new research infrastructures and strategies. To address these challenges, an interdisciplinary team of scientists from the Karlsruhe Institute of Technology (KIT) and the German Meteorological Service (DWD) created the TEstbed for Exploring Machine LEarning in Atmospheric Prediction (TEEMLEAP). Implemented on KIT's supercomputer HoreKa, the TEEMLEAP testbed simulates the entire operational weather forecasting chain using ERA5 reanalysis data as pseudo-observations and DWD's Basic Cycling environment for conducting assimilation-prediction-cycling experiments. Moreover, first steps are taken toward the integration of new data-driven components like FourCastNet and ML-based post-processing methods. The TEEMLEAP testbed allows systematic investigation of a wide range of issues related to weather forecasting such as optimizing the observational system, uncertainty quantification, and developing hybrid systems that integrate ML with physics-based models. This document outlines the testbed's setup, demonstrates its functionality with a pilot experiment, and discusses examples of potential applications. Future plans include creating educational modules and developing a higher-resolution regional version of the testbed that could be used for assimilating field campaign observations.