Earth system models: How AI and satellites should improve climate predictions

Artificial intelligence (AI) should make it easier to predict the future course of climate change in the future. A research team from Germany, Spain and the USA, led by Veronika Eyring from the German Aerospace Center (DLR), has developed a new approach to integrating machine learning into models of the Earth system. The technology is designed to learn the influence of a specific atmospheric process on the basis of an existing climate model with a very high spatial resolution in the kilometer range, for example, and to use this new knowledge in the "coarse-meshed" Earth system model, thereby making it significantly more efficient. The evaluation of satellite-based earth observation data, which is considered an important basis for climate and environmental research, plays a key role here.

Earth system models such as Earth4All take into account important processes in the atmosphere and its interactions with other components such as oceans and land. In principle, they can therefore provide far-reaching predictions of the entire system of the blue planet. This requires the processing of large amounts of data, which are limited in terms of computing time and spatial resolution. This results in inaccuracies and systematic errors. Avoiding these as far as possible is one of the greatest challenges in climate modeling. Eyring and her team use machine learning methods to improve the representation of processes in the simulations that cannot be explicitly resolved in the models. These are of central importance for climate dynamics.

Predictions with "unprecedented accuracy"

Alternative simulations with high-resolution climate models in the kilometer range achieve higher accuracy compared to observational data. However, they are only suitable for climate predictions of several decades to a limited extent, as the calculations are data-intensive and therefore very costly. The new approach, on which the scientists involved have just published an article in the journal Nature Geoscience, combines models across different scales and different process complexity with the systematic use of AI and satellite data. The latter traditionally help to assess and evaluate climate and Earth system models, which are then used for climate predictions and the derivation of action measures for sectors such as energy, aviation and transport.

AI-supported climate modeling will also form the basis for more realistic digital twins of the Earth system that are scalable, user-interactive and adaptable. Such computer simulations help to better understand the interplay of movements and forces in the physical world. Basically, the important resource of satellite-based Earth observation data needs to be used much more intensively "to calibrate, evaluate and improve global forecasting models," emphasizes Eyring from the DLR Institute of Atmospheric Physics in Oberpfaffenhofen. She is certain that the combination with AI "will enable us to predict the complexity of the Earth's future climate and extreme events with unprecedented accuracy". Back in August, the professor and her colleagues published a more general study on pushing the boundaries of climate modeling with the help of machine learning in the journal Nature Climate Change.

(mma)