A New Generation of Global Climate Models Augmented by AI
, Professor in Mathematics and Atmosphere/Ocean Science, New York University
Climate simulations have long been invaluable in understanding and predicting global and regional climate change. Their fidelity has been limited by computing capabilities, leading to inaccurate parametrizations of key unresolved processes such as convection, cloud, or mixing, and consequently to biases in large-scale phenomena such as temperature, rainfall, and sea level. These unresolved processes have posed a significant hurdle in enhancing climate simulations and their predictions.
A promising paradigm shift is now underway, fueled by the explosion of climate data and the formidable potential of machine learning (ML) algorithms to parametrize subgrid processes in climate models. We'll present a suite of global simulations in which ML parameterizations replace, or augment, ad-hoc representations of ocean and sea-ice subgrid processes. The simulations are performed with the NOAA-Geophysical Fluid Dynamics Laboratory’s MOM6-based global climate models OM4 and CM4. We'll discuss how the data-driven parameterizations of ocean mixing and sea-ice affect large-scale biases and variability of relevant climate fields, and associated mechanisms. This new generation of data-informed simulations has the potential to provide more reliable climate projections at global and local scales. This is collaborative work as part of M2LInES: https://m2lines.github.io/