What Might We Expect Over the Next Decade for Machine Learning in Weather and Climate Modeling?
As part of the Machine Learning for Earth System Modeling training course, I had the pleasure of moderating a panel discussion on “What Might We Expect Over the Next Decade for Machine Learning in Weather and Climate Modeling?” The discussion brought together perspectives from ECMWF, Predictia, the University of Lausanne, and KIT, with contributions spanning data-driven forecasting, climate emulation, Earth system modeling, high-performance computing, data infrastructure, and university research.
The discussion started from a striking observation: only a few years ago, there was still considerable skepticism about whether machine learning models could compete with numerical weather prediction systems, especially at medium-range timescales. Since then, data-driven models have progressed rapidly. The question has therefore shifted from whether machine learning can contribute to weather and climate modeling to how it will reshape the way we build models, use data, evaluate forecasts, and organize scientific workflows.
One important theme was the future architecture of machine-learning-based Earth system models. Some panelists highlighted the potential of large, flexible foundation models with shared latent representations, in which interactions among atmosphere, ocean, sea ice, waves, and other components are learned directly from data. Others emphasized the value of more modular ecosystems, where specialized models, hybrid systems, and orchestration layers can be combined as needed for each task, uncertainty, and validity domain. This tension between unified models and modular model families may become one of the defining questions of the next decade.
The role of data was another central point. Machine learning models are only as useful as the data ecosystems that support them. The discussion highlighted the need for AI-ready observational datasets, better access to campaign data, harmonized formats, clearer metadata, and workflows that make it easier to combine reanalyses, simulations, satellite observations, and field measurements. In this sense, progress in machine learning for weather and climate will depend not only on new architectures but also on the quality, accessibility, and usability of the data infrastructure around them.
A particularly interesting part of the discussion concerned how scientific work itself may change. Machine learning is not only producing new forecast models; it is also changing coding practices, software development, verification, data compression, visualization, and interactive use of models. Instead of simply downloading fixed-forecast products, future users may increasingly interact directly with flexible models, refine forecasts on demand, or explore “what-if” scenarios more dynamically.
The panel also raised an important question for universities and early-career scientists: what should a weather and climate scientist know in 2035? A deep physical understanding will remain essential, but it may need to be combined with AI literacy, data skills, awareness of uncertainty, and the ability to interrogate models critically. Universities may not always compete with large weather centers or technology companies in training the largest models. Still, they can play a crucial role in understanding processes, developing smaller, targeted models, probing physical consistency, and educating the next generation of scientists.
From my perspective as a moderator, the discussion reinforced that the next decade will not be about building only more accurate models. It will also be about building more useful, transparent, and scientifically meaningful systems. Machine learning may help us generate forecasts faster, emulate expensive simulations, downscale information, and discover new relationships in complex data. But it also forces us to rethink trust, evaluation, robustness, and the role of human expertise in increasingly automated workflows.
The future of machine learning in weather and climate modeling will likely be plural rather than singular. It may involve foundation models, hybrid models, regional downscaling systems, climate emulators, AI-assisted interpretation tools, and physical models that remain essential for data generation, extrapolation, and scientific grounding. The real challenge will be to connect these pieces into workflows that are not only technically powerful but also reliable, interpretable, and useful for science and society.
A recording of the panel discussion is available on YouTube: