NCAR Community Research Earth Digital Intelligence Twin (CREDIT) (Core Science Keynote)

Artificial Intelligence (AI) Numerical Weather Prediction (NWP) models have advanced rapidly in their capabilities and have shown promising performance with a variety of architectures, especially transformers and graph neural networks. However, the first generation of AI NWP models also exhibit some significant prediction issues, including overly smooth predictions, patch artifacts, and performance degradations when running further away from the training date period. These issues raise many questions about the sensitivities to choices regarding data pre-processing, model training, and how the models are iterated forward in time that should be explored more thoroughly by the weather and climate community. While some of the AI NWP groups have open sourced portions of their training and inference pipelines, existing open source platforms require substantial machine learning and programming expertise to train new models and perform sensitivity analyses.

To address this need, the Machine Integration and Learning for Earth Systems (MILES) group at NSF NCAR has developed the Community Runnable Earth Digital Intelligence Twin (CREDIT) framework, a user-friendly open research platform for training, running, and analyzing AI NWP systems. CREDIT supports the processing of multiple global and regional datasets, including ERA5 and CONUS404 and in particular has support for processing hourly ERA5 model level output. Using PyTorch as the deep learning backend, we have trained both existing architectures and new transformer and graph architectures with the framework and have achieved stable hourly forward predictions with a single model out to two weeks at 0.28 degree spacing and one year at 1 degree spacing. The presentation showcases analyses of model performance, comparisons of 1-hour and 6-hour time stepping, and sensitivities to choice of initialization time and normalization scheme. Finally, opportunities to contribute to CREDIT will be discussed.