12.01.2026: Ocean Circulation and Climate Dynamics Colloquium
Lorenzo Zampieri (ECMWF), Bonn: "Machine Learning Beyond the Atmosphere: Towards Data-Driven Earth System Models at ECMWF"
When? Monday 12 January 2026 at at 11 am
Where? Conference Room 5-1.214, Wischhofstr. 1-3 and online
via Meeting link: https://geomar.webex.com/geomar-en/j.php?MTID=ma73a9dd4db0f5cb6275b471e1c5e6005
Meeting number: 2789 008 5026
Password: nhPGUfGj553
Abstract:
Machine learning (ML) has recently demonstrated remarkable skill in atmospheric weather prediction, raising the prospect of a new class of data-driven Earth system models. At ECMWF, this effort is embodied in the Artificial Intelligence Forecasting System (AIFS), which provides operational machine learning–based forecasts alongside traditional numerical prediction systems. While early successes have focused on the atmosphere, extending ML approaches to the ocean and to coupled Earth system modelling poses distinct scientific and technical challenges. In this talk, I will first introduce the AIFS framework, outlining how single and ensemble forecasts are produced and how data-driven models differ conceptually from physics-based approaches. I will then move beyond the atmosphere to discuss the extension of AIFS towards a coupled Earth system, with a particular emphasis on marine components. This includes machine learning representations of sea ice, ocean waves, and the three-dimensional ocean, trained on reanalysis data and designed to operate across a range of spatial and temporal scales. Special attention will be given to challenges specific to ocean modelling, such as the slower evolution of deep-ocean variables, data sparsity at depth, and the need to maintain physical consistency over longer timescales. I will describe strategies developed to address these issues, including alternative training targets, scaling approaches for tendencies, and methods for handling missing values and enforcing physical bounds. Finally, I will discuss different strategies for coupling machine learning components in Earth system models, ranging from loosely coupled component models to more tightly integrated approaches, and compare these with established coupling practices in numerical models. Overall, the talk aims to provide an accessible overview of the current state of data-driven ocean and coupled Earth system modelling, highlighting both recent progress and key open questions relevant to marine and climate research.
