Machine Learning

Machine learning (ML) has driven breakthroughs across diverse fields - from computer vision to generating human-like text with large language models (LLMs), and more recently, advancing climate science. ML is emerging as an important tool for predictive modeling in weather prediction, and for improving our understanding of climate physics. Key advantages of ML in this field are the study of new or more detailed physical insights through knowledge or model discovery, as well as faster and more computationally efficient accurate weather forecasting.

In the Atmospheric Physics group at AWI, we are exploring the potential of ML to enhance Arctic climate research by correcting biases in atmospheric datasets. Bias correction involves reducing systematic errors present in reanalysis datasets, which are extensively used in climate research but less reliable than direct observations of climate variables. This process is especially crucial in the Arctic, where the harsh environment limits observational coverage, necessitating wide-ranging data assimilation techniques to fill in the gaps. As a result, developing more accurate and comprehensive Arctic atmospheric datasets is critical for understanding the region’s rapidly changing climate. 

One ongoing project involves an analysis of using a Multi-Layer Perceptron (MLP) neural network to bias correction for multiple surface variables, with the goal of improving surface energy budget estimates. In this study, an MLP was trained to correct ERA5 data to predict surface fluxes observed during several key Arctic expeditions—MOSAiC, SHEBA, AO2018, and ARTofMELT. The results show a reduction of up to 50% in the root mean square error of surface energy budget estimates compared to ERA5, paving the way for producing ML based  bias-corrected estimates of surface energy fluxes over Arctic sea ice.

In a complementary effort, we are exploring a similar objective, but instead fine-tuning a pre-trained foundational Transformer-based model to correct biases in near-surface Arctic temperature estimates. Foundation models are deep learning systems trained on vast and diverse datasets, enabling them to generalise across a wide range of tasks. By adapting such a pre-trained model for Arctic bias correction, we aim to leverage its learned representations of global climate dynamics to more accurately capture near-surface temperature patterns in the Arctic with very sparse observational data. This project lays the groundwork for a foundational ML model-based framework that complements traditional reanalysis-driven data assimilation.