Model mixes AI and physics to do global forecasts

Right now, the world’s best weather forecast model is a General Circulation Model, or GCM, put together by the European Center for Medium-Range Weather Forecasts. A GCM is in part based on code that calculates the physics of various atmospheric processes that we understand well. For a lot of the rest, GCMs rely on what’s termed “parameterization,” which attempts to use empirically determined relationships to approximate what’s going on with processes where we don’t fully understand the physics.

Lately, GCMs have faced some competition from machine-learning techniques, which train AI systems to recognize patterns in meteorological data and use those to predict the conditions that will result over the next few days. Their forecasts, however, tend to get a bit vague after more than a few days and can’t deal with the sort of long-term factors that need to be considered when GCMs are used to study climate change.

On Monday, a team from Google’s AI group and the European Centre for Medium-Range Weather Forecasts are announcing NeuralGCM, a system that mixes physics-based atmospheric circulation with AI parameterization of other meteorological influences. Neural GCM is computationally efficient and performs very well in weather forecast benchmarks. Strikingly, it can also produce reasonable-looking output for runs that cover decades, potentially allowing it to address some climate-relevant questions. While it can’t handle a lot of what we use climate models for, there are some obvious routes for potential improvements.

Meet NeuralGCM

NeuralGCM is a two-part system. There’s what the researchers term a “dynamical core,” which handles the physics of large-scale atmospheric convection and takes into account basic physics like gravity and thermodynamics. Everything else is handled by the AI portion. “It’s everything that’s not in the equations of fluid dynamics,” said Google’s Stephan Hoyer. “So that means clouds, rainfall, solar radiation, drag across the surface of the Earth—also all the residual terms in the equations that happen below the grid scale of about roughly 100 kilometers or so.” It’s what you might call a monolithic AI. Rather than training individual modules that handle a single process, such as cloud formation, the AI portion is trained to deal with everything at once.

Critically, the whole system is trained concurrently rather than training the AI separately from the physics core. Initially, performance evaluations and updates to the neural network were performed at six-hour intervals since the system isn’t very stable until at least partially trained. Over time, those are stretched out to five days.

The result is a system that’s competitive with the best available for forecasts running out to 10 days, often exceeding the competition depending on the precise measure used (in addition to weather forecasting benchmarks, the researchers looked at features like tropical cyclones, atmospheric rivers, and the Intertropical Convergence Zone). On the longer forecasts, it tended to produce features that were less blurry than those made by pure AI forecasters, even though it was operating at a lower resolution than they were. This lower resolution means larger grid squares—the surface of the Earth is divided up into individual squares for computational purposes—than most other models, which cuts down significantly on its computing requirements.

Despite its success with weather, there were a couple of major caveats. One is that NeuralGCM tended to underestimate extreme events occurring in the tropics. The second is that it doesn’t actually model precipitation; instead, it calculates the balance between evaporation and precipitation.

But it also comes with some specific advantages over some other short-term forecast models, key among them being that it isn’t actually limited to running over the short term. The researchers let it run for up to two years, and it successfully reproduced a reasonable-looking seasonal cycle, including large-scale features of the atmospheric circulation. Other long-duration runs show that it can produce appropriate counts of tropical cyclones, which go on to follow trajectories that reflect patterns seen in the real world.