× off
Improving climate models and forecasts by learning from observed and simulated data. To improve climate models, model components encoding domain-specific knowledge must learn from various climate statistics obtained from Earth observations or regional high-resolution simulations. Ideally, model components learn together, and by quantifying their joint uncertainty, through a shared layer of data assimilation and machine learning tools that wraps all model components to reveal and reduce compensating errors between components.5. These model calibrations and uncertainty quantifications require large ensembles of climate simulations, and large ensembles are also needed to sample the space of potential climate effects.2. These simulation ensembles can be generated at moderately high resolution (10–50 km), but not yet at the kilometer scale. Credit: Nature climate change (2023). DOI: 10.1038/s41558-023-01769-3
To date, climate models have been challenged to provide high-resolution forecasts – with quantified uncertainty – needed by a growing number of adaptation planners, from local decision-makers to the private sector, who require detailed assessments of climate risks. Local face.
This calls for a step change in the accuracy and utility of climate forecasts, according to the authors of the paper “Harnessing AI and Computing to Advance Climate Modeling and Prediction” through artificial intelligence.
The comment was published in Nature climate change by a group of international climate scientists including CMCC Scientific Director Giulio Boccaletti and CMCC President Antonio Navara.
A proposed approach for a step change in climate modeling is to focus on global models with 1-km horizontal resolution. However, the authors explain, although kilometer-scale models have been referred to as Earth’s “digital twins,” they still have the same limitations and biases as existing models. Moreover, given the high computational cost, they impose limitations on the size of simulation ensembles, which are necessary both to calibrate the inevitable empirical models of unresolved processes and to quantify uncertainty.
Overall, kilometer-scale models do not offer a step change in accuracy that would accept the limitations they impose.
Instead of prioritizing kilometer-scale resolution, the authors propose a balanced approach focusing on creating large ensembles of moderately high resolution (10–50 km, up from about 100 km, which is standard today) simulations that take advantage of advances in computing and AI learning. from the data.
By modestly increasing global resolution while making extensive use of observational and simulated data, this approach is more likely to achieve the goals of climate modeling for risk assessment, including reducing model error and quantifying uncertainty, and enables widespread adoption.
1,000 simulations at 10-km resolution cost the same as 1 simulation at 1-km resolution. “As computer performance increases we must push the resolution frontier further, yet climate modeling will need to focus on resolution in the 10-50 km range over the next decade,” the authors write. “Importantly, climate models must be developed so that they can be used and improved globally in a comprehensive and distributed research program through rapid iterations that do not concentrate resources in a few discrete centers that would be required if the focus were on kilometer-scale global modeling.”
More information:
Tapio Schneider et al, Harnessing AI and Computing to Advance Climate Modeling and Prediction, Nature climate change (2023). DOI: 10.1038/s41558-023-01769-3
Journal Information:
Nature climate change
Provided by the CMCC Foundation – Euro-Mediterranean Center on Climate Change