Currently, the best weather forecast models in the world are the General Circulation Models (GCMs) put together by the European Centre for Medium-Range Weather Forecasts. GCMs are based in part on codes that calculate the physics of various atmospheric processes that we understand well. For much of the rest, GCMs rely on what’s called “parameterization,” which uses empirically determined relationships to try to approximate what’s going on in processes whose physics we don’t fully understand.
Recently, GCMs have faced competition from machine learning techniques, which train AI systems to recognize patterns in weather data and use them to predict weather conditions over the next few days. But the predictions tend to be a bit vague beyond a few days and don’t address the longer-term factors that need to be considered when using GCMs to study climate change.
On Monday, a team from Google’s AI group and the European Centre for Medium-Range Weather Forecasts announced NeuralGCM, a system that combines physics-based atmospheric circulation with AI parameterization of other weather influences. Neural GCM is computationally efficient and performs very well in weather forecasting benchmarks. Surprisingly, it can produce reasonable outputs even over multi-decade runs, potentially addressing several climate-related problems. While it cannot handle many of the applications for which climate models are used, there is clearly room for improvement.
About NeuralGCM
NeuralGCM is a two-part system. There’s what researchers call a “dynamical core,” which handles the physics of large-scale atmospheric convection and takes into account fundamental physics like gravity and thermodynamics. Everything else is handled by the AI part. “It’s everything that’s not in the fluid dynamics equations,” says Google’s Stephan Hoyer. “So clouds, precipitation, solar radiation, drag on the Earth’s surface, and all the residual terms in the equations that happen below a grid scale of about 100 kilometers or so.” This is what you might call a monolithic AI: instead of training separate modules to handle a single process, like cloud formation, the AI part is trained to handle everything at once.
The key is to train the entire system at the same time, rather than training the AI from each physical core individually. Initially, performance evaluations and neural network updates were performed at six-hour intervals, as the system is not very stable until it is at least partially trained. Over time, they were extended to five days.
The result was a system that could compete with the best systems for 10-day forecasts, and often outperformed them depending on the exact measurements used. (In addition to weather benchmarks, the researchers also looked at features such as tropical cyclones, atmospheric rivers, and intertropical convergence zones.) For longer-term forecasts, the system tended to produce less blurred features than those produced by pure AI forecasters, even though it operated at a lower resolution than those produced by pure AI forecasters. This lower resolution meant that the grid squares (the Earth’s surface is divided into individual squares for calculations) were larger than in most other models, which significantly reduced computational requirements.
Despite its success in weather forecasting, NeuralGCMs come with some big caveats: first, they tend to underestimate extreme events that occur in the tropics, and second, they don’t actually model precipitation, but rather calculate a balance between evaporation and precipitation.
But the model has several advantages over other short-term forecast models, the main one being that it’s not limited to short-term runs. Researchers have run the model for up to two years, and it does a good job of reproducing plausible seasonal cycles, including large-scale features of atmospheric circulation. Other longer runs have shown that it can generate a good number of tropical cyclones, which follow trajectories that reflect patterns seen in the real world.