October 5, 2022 — Computers are already using artificial intelligence to improve the resolution of blurry images, to create images that mimic the style of particular painters from photographs, and to render realistic portraits of people who don’t actually exist. The underlying method is based on so-called GANs (Generative Adversarial Networks).
A team led by Niklas BoerProfessor of Earth System Modeling at Technical University of Munich (TUM) and researcher at the Potsdam Institute for Climate Impact Research (PIK) is now applying these machine learning algorithms to climate research. The research group recently published their findings in the journal “Intelligence of natural machines“.
Not all processes can be taken into account
“Climate models differ from models used to make weather forecasts, particularly with regard to their larger time horizon. The forecast horizon of weather forecasts is days, while climate models run simulations over decades or even centuries,” explained Philip Hess, lead author of the study and research associate at the TUM Chair for Earth System Modelling. The weather can be predicted quite accurately for a few days; the prediction can then be verified based on actual observations. When it comes to climate, however, the goal is not a weather-based prediction, but among other things projections of the impact of increasing greenhouse gas emissions on Earth’s climate. long-term.
However, climate models still cannot fully account for all relevant climate processes. This is partly because certain processes are not yet sufficiently understood, and partly because detailed simulations would take too long and require too much computing power.
“As a result, climate models still cannot represent extreme precipitation events as we would like. Therefore, we started using GANs to optimize these models with respect to their precipitation production,” Boers said.
Optimizing climate models with meteorological data
Basically, a GAN consists of two neural networks. One network tries to create an example from a previously defined product, while the other tries to distinguish this artificially generated example from real examples. The two networks thus compete with each other, improving constantly. A practical application of GANs would be to “translate” landscape paintings into realistic photographs. The two neural networks take photorealistic images generated on the basis of the paint and send them back and forth until the images created can no longer be distinguished from actual photographs.
Niklas Boers’ team took a similar approach: the researchers used a relatively simple climate model to demonstrate the potential of machine learning to improve these models. The team’s algorithms use observed weather data. Using this data, the team trained GAN to modify the climate model simulations so that they could no longer be distinguished from actual weather observations.
“This way, the degree of detail and realism can be increased without the need for complicated additional process calculations,” said Markus Druckeclimate modeler at PIK and co-author of the study.
GANs can reduce electricity consumption in climate modeling
Even relatively simple climate models are complex and are processed using supercomputers that consume large amounts of power. The more detail the model takes into account, the more complicated the calculations become and the greater the amount of electricity used. The computations involved in applying a trained GAN to a climate simulation, however, are negligible compared to the amount of computation required for the climate model itself.
“Using GANs to make climate models more detailed and realistic is therefore practical not only for improving and speeding up simulations, but also in terms of saving electricity,” Hess said.
Source: Technische Universität München
