Currently, the best global climate model is the General Circulation Model, or GCM, compiled by the European Center for Medium-Range Weather Forecasts. The GCM component is based on code that calculates the physics of various atmospheric processes that we understand well. For the most part, GCMs rely on what’s called “parameterization,” which tries to use objective relationships to quantify what’s happening in processes where we don’t fully understand the physics.
Recently, GCMs have faced some competition from machine learning methods, which train AI systems to recognize patterns in weather data and use those to predict conditions that will end in a few days. the following. However, their predictions are somewhat imprecise after a few days and cannot handle the long-term trends that must be considered when GCMs are used to study climate change. of heaven.
On Monday, a team from Google’s AI team and the European Center for Medium-Range Weather Forecasts announced NeuralGCM, a system that combines physics-based atmospheric circulation with AI parameterization of other weather influences. Neural GCM performs well compared to climate models. Remarkably, it can also generate a reasonable output for games that span decades, which would allow it to answer questions related to climate. Although it can’t handle many of the things we use for climate models, there are clear lines of potential improvement.
Meet NeuralGCM
NeuralGCM is a two-part system. There is what the researchers call a “dynamical core,” which deals with the physics of large-scale atmospheric convection and takes into account fundamental physics such as gravity and thermodynamics. Everything else is handled by the AI component. “It’s everything that’s not in liquid energy,” said Google’s Stephan Hoyer. “So that means clouds, rain, sunlight, drag on the Earth—and all the other terms in the equations that happen below a network of about 100 kilometers or so. .” It is what you would call monolithic AI. Instead of training individual modules that handle a single function, such as cloud formation, the AI component is trained to handle everything at once— once.
Ideally, the entire system is trained simultaneously instead of training the AI separately from the physics base. Initially, performance evaluations and improvements to the neural network were performed in six-hour intervals as the system is unstable until it is at least partially trained. Over time, they are stretched to five days.
The result is a system that competes with those available for forecasts beyond 10 days, often surpassing the competition in terms of the precision used (besides meteorological standards, researchers are looking at features such as tropical cyclones, rivers of atmosphere, and the Intertropical Confluence). In long-term displays, it tended to produce features that were more blurry than those created by pure AI wizards, even though it was running at a lower resolution than it was. This lower resolution means a larger grid area – the Earth’s surface is divided into individual squares for computing purposes – than many other models, which greatly reduces its computing requirements.
Despite the weather’s success, there were a few big caveats. One is that NeuralGCM tended to underestimate extreme events in the tropics. The second is that it doesn’t really model rain; rather, it measures the balance between evaporation and precipitation.
But it also comes with some advantages over other types of short-term forecasting, the most important of which is that it is not necessarily limited to the short term. The researchers let it run for up to two years, and it succeeded in producing a reasonable seasonal cycle, including large parts of the atmospheric cycle. Some long-term measurements show that they can produce accurate numbers of tropical cyclones, which continue to follow tracks that reflect patterns seen in the real world.
#model #combines #physics #global #predictions
