Computers are already using artificial intelligence to enhance fuzzy image resolution. The fundamental method depends on the so-called GANs (Generative Adversarial Networks).
Machine Learning Algorithms" />
Machine learning gives researchers a more detailed view of our global climate. Image Credit: Vladi333/Shutterstock.com
A group headed by Niklas Boers, Professor for Earth System Modelling at the Technical University of Munich (TUM) and Researcher at the Potsdam Institute for Climate Impact Research (PIK) is now using these machine learning algorithms to research climate. The study team recently published its results in the Nature Machine Intelligence journal.
Not All Processes Can Be Taken into Account
Climate models differ from the models used to make weather forecasts, especially in terms of their broader time horizon. The forecast horizon for weather predictions is several days, while climate models perform simulations over decades or even centuries.
Philipp Hess, Study Lead Author and Research Associate, TUM Professorship for Earth System Modelling, Technical University of Munich
Weather can be accurately forecasted for a few days; the forecast can then be assessed depending on actual observations. However, the goal is not a time-based forecast when the climate is concerned, but how increasing greenhouse gas emissions will affect the climate of the Earth in the future—among projections of other things.
Climate models, however, still cannot consider all appropriate climate processes. This is because some methods are not accurately understood and elaborate simulations need much computing power.
As a result, climate models still can’t represent extreme precipitation events the way we’d like. Therefore, we started using GANs to optimize these models with regard to their precipitation output.
Niklas Boers, Professor and Researcher, Earth System Modeling, Technical University of Munich
Optimizing Climate Models with Weather Data
A GAN comprises two neural networks. One network tries to form an example from an already defined product, while the second network attempts to differentiate this artificially created example from real examples. The two networks contend with each other, seamlessely enhancing the process.
“Translating” landscape paintings into real-life photos is one real-time application of GANs. The two neural networks capture photo-realistic pictures produced based on the painting and deliver them back and forth until the images generated cannot be differentiated from real photos.
The team of Niklas Boers took a similar approach: The scientists employed a relatively easier climate model to establish the capacity of using machine learning to enhance these models. The algorithms of the team used observed weather data. The team trained the GAN using this data to modify the climate model’s simulations so that they could no longer be differentiated from actual weather observations.
This way the degree of detail and realism can be increased without the need for complicated additional process calculations.
Markus Drücke, Study Co-Author and Climate Modeler, Potsdam Institute for Climate Impact Research
GANs Can Reduce Electricity Consumed in Climate Modeling
Comparatively easier climate models are complicated and processed with supercomputers that take up large quantities of energy. The calculations get more complicated and the amount of electricity consumed increases with the more details the model considers.
However, compared to the amount of calculation needed for the climate model, the calculations employed in using a trained GAN for a climate simulation are insignificant. “Using GANs to make climate models more detailed and more realistic is thus practical not only for the improvement and acceleration of the simulations, but also in terms of saving electricity,” Philipp Hess concludes.
Journal Reference
Hess, P., et al. (2022). Physically constrained generative adversarial networks for improving precipitation fields from Earth system models. Nature Machine Intelligence. doi.org/10.1038/s42256-022-00540-1.