Adapting Machine Learning Forecasting Models more efficiently with transfer learning

Adapting Machine Learning Forecasting Models more efficiently with transfer learning

Victor Marchais

Joint Center for Satellite Data Assimilation / GMAO

In recent years, the development of trained machine learning-based forecasting models has gained significant attention within the meteorological community due to their advantages, including rapid processing speeds, strong performance metrics, and the ability to automatically handle linearization tasks. However, these models are often tailored to one specific grid and set of variables, typically trained on pressure levels and ERA5 reanalysis data. Compared to regular models it makes them less flexible when applied to different contexts and models (e.g. UFS, GEOS, MPAS). Training such models from scratch is resource-intensive and logistically challenging. In this study, we explore various transfer learning methodologies as a cost-effective and efficient alternative for model generation.

We experiment with several transfer learning methods on the hybrid model Neural GCM. We changed the initial training data for NASA’s model GEOS’ grid and variables. We demonstrate that transfer learning not only accelerates the training process but also enhances the overall performance of the forecasting models compared to models developed from the ground up. These findings suggest that transfer learning can be a viable strategy for extending the applicability of machine learning models across diverse meteorological datasets and tasks, including data assimilation.

Keywords: Machine Learning, neural network, transfer learning, General Circulation Model, GEOS