AI speeds up climate computations

September 10, 2018, Ludwig Maximilian University of Munich
AI speeds up climate computations
General circulation models typically simulate the global behavior of the atmosphere on grids whose cells have dimensions of around 50 km. But many clouds extend over distances of only a few hundred meters, much smaller than the grid cells typically used in simulations – and they are highly dynamic. Both features make them extremely difficult to model realistically. Credit: Robert Kneschke / fotolia

Realistic climate simulations require huge reserves of computational power. An LMU study now shows that new algorithms allow interactions in the atmosphere to be modeled more rapidly without loss of reliability.

Forecasting global and local climates requires the construction and testing of mathematical . Since such models must incorporate a plethora of physical processes and interactions, climate simulations require enormous amounts of . And even the best models inevitably have limitations, since the phenomena involved can never be modeled in sufficient detail. In a project carried out in the context of the DFG-funded Collaborative Research Center "Waves to Weather", Stephan Rasp of the Institute of Theoretical Meteorology at LMU (Director: Professor George Craig) has now looked at the question of whether the application of can improve the efficacy of climate modelling. The study, which was performed in collaboration with Professor Mike Pritchard of the University of California at Irvine und Pierre Gentine of Columbia University in New York, appears in the journal PNAS.

General circulation models typically simulate the global behavior of the atmosphere on grids whose cells have dimensions of around 50 km. Even using state-of-the-art supercomputers the relevant that take place in the atmosphere are simply too complex to be modelled at the necessary level of detail. One prominent example concerns the modelling of clouds which have a crucial influence on climate. They transport heat and moisture, produce precipitation, as well as absorb and reflect solar radiation, for instance. Many clouds extend over distances of only a few hundred meters, much smaller than the grid cells typically used in simulations – and they are highly dynamic. Both features make them extremely difficult to model realistically. Hence today's models lack at least one vital ingredient, and in this respect, only provide an approximate description of the Earth system.

In the new study, Rasp and his co-authors have used a form of machine learning known as neural networks to adaptively tune an designed to capture cloud behavior. The algorithm was trained using data obtained from high-resolution simulations that explicitly included heat transfer of clouds. "Following the training period, the algorithm was not only able to reproduce the results obtained with the fine-scale, cloud-resolving model, but did so much more efficiently," says Stephan Rasp. According to George Craig, "the study shows that the method has the potential to describe complex phenomena in greater detail and therefore promises to enhance the quality of ."

In this initial test, the authors used an idealized model to test the feasibility of the approach with reduced complexity. In the next step, Rasp, Pritchard and Gentine plan to train the algorithm with real data.

Explore further: Machine learning may be a game-changer for climate prediction

More information: Stephan Rasp et al. Deep learning to represent subgrid processes in climate models, Proceedings of the National Academy of Sciences (2018). DOI: 10.1073/pnas.1810286115

Related Stories

The complexities of clouds and the seeds that make them

January 12, 2018

Clouds are complicated. Each cloud formation depends on the timing of the water cycle, in which water evaporates from Earth's surface, condensates in the atmosphere and falls back down, as well as the types of aerosols in ...

Recommended for you

2 comments

Adjust slider to filter visible comments by rank

Display comments: newest first

jaimesald
5 / 5 (1) Sep 10, 2018
Looks like this might be suitable for Weather Prediction.
howhot3
not rated yet Sep 11, 2018
What a great use for future tech,

Please sign in to add a comment. Registration is free, and takes less than a minute. Read more

Click here to reset your password.
Sign in to get notified via email when new comments are made.