Currently, the best weather forecast model in the world is the General Circulation Model (GCM), established by the European Center for Medium-Range Weather Forecasts. The GCM is based in part on code that calculates the physics of various atmospheric processes that we are familiar with. For much of the rest, GCMs rely on so-called “parameterization,” which attempts to use empirically determined relationships to approximate situations in which we do not fully understand physical processes.
Recently, GCM has faced some competition from machine learning technologies, which train artificial intelligence systems to recognize patterns in weather data and use those patterns to predict conditions that will occur in the coming days. However, their predictions tend to become a bit fuzzy after a few days and cannot handle the long-term factors that need to be considered when using GCMs to study climate change.
On Monday, teams from the Google Artificial Intelligence Group and the European Center for Medium-Range Weather Forecasts announced the launch of NeuralGCM, a system that combines physics-based artificial intelligence parameterization of atmospheric circulation with other meteorological influences. Neural GCM is computationally efficient and performs well on weather forecasting benchmarks. Remarkably, it can also produce reasonable-looking outputs over decades of operation, potentially solving some climate-related problems. While it can't handle much of what we use climate models for, there are some obvious potential avenues for improvement.
Meet NeuralGCM
NeuralGCM is a two-part system. The researchers call it the “dynamic core,” which deals with the physics of large-scale atmospheric convection and takes into account fundamental physical principles such as gravity and thermodynamics. Everything else is handled partially by the artificial intelligence. “That’s everything outside of the fluid dynamics equation,” said Google’s Stephan Hoyer. “So that means clouds, rainfall, solar radiation, drag on the Earth’s surface, and what happens in the equation at about 100 kilometers All remaining items below the grid scale of left and right.” This is what you call overall artificial intelligence. Rather than training a single module to handle a single process (such as cloud formation), the AI part trains it to handle everything at once.
Crucially, the entire system is trained simultaneously, rather than training the AI separately from the physical core. Initially, the neural network's performance was evaluated and updated every six hours because the system was not very stable until it had been at least partially trained. Over time, these periods extend to five days.
The result is a system that rivals the best forecasts up to 10 days out and often outperforms competitors based on the precise measurements used (in addition to weather forecasting benchmarks, the researchers looked at tropical cyclones, atmospheric rivers, and atmospheric layers, among others feature). Over longer forecasts, it tends to produce sharper features than pure AI predictors, although it operates at a lower resolution than pure AI predictors. This lower resolution means larger grid squares than most other models (the Earth's surface is divided into individual squares for computational purposes), which greatly reduces its computational requirements.
Despite its success with the weather, there were some major caveats. One is that NeuralGCM tends to underestimate extreme events occurring in the tropics. The second is that it doesn't actually simulate precipitation; Instead, it calculates the balance between evaporation and precipitation.
But it also has some specific advantages over some other short-term forecasting models, the key one being that it's not actually limited to short-term runs. The researchers let it operate for up to two years, and it successfully reproduced what appeared to be reasonable seasonal cycles, including large-scale features of atmospheric circulation. Other long-term runs have shown that it can produce adequate counts of tropical cyclones that continue to follow trajectories that mirror patterns seen in the real world.
