How Google's NeuralGCM overcomes limitations of traditional weather forecasting models
Google in partnership with the European Centre for Medium-Range Weather Forecasts (ECMWF), has introduced a machine learning-based weather forecasting method called NeuralGCM. This innovative approach merges traditional physics-based modeling with ML, to enhance both the accuracy and efficiency of atmospheric simulations. According to Google, NeuralGCM can generate weather forecasts spanning 2-15 days, that surpass the current gold-standard physics-based model in terms of accuracy.
Physics and AI work together
Google has stated that while NeuralGCM has not yet been integrated into a full climate model, it represents a significant advancement toward creating more powerful and accessible ones. The new method combines the strengths of ML techniques with general circulation models, which have been the cornerstone of weather prediction for the past half-century. Stephan Hoyer, an AI researcher at Google Research, emphasized that it's not about physics versus AI but rather physics and AI working together.
A blend of conventional models and AI
The NeuralGCM approach utilizes a conventional model to calculate large atmospheric changes necessary for a prediction. It then incorporates AI to rectify errors that accumulate on smaller scales, typically for predictions on scales less than 25km. Hoyer explained, "That's where we inject AI very selectively to correct the errors that accumulate on small scales."
Overcoming limitations
Traditional climate models often generate errors due to an incomplete understanding of Earth's climate and the way these models are constructed. They divide the globe into cubes, typically 50-100km on each side, extending from the surface up into the atmosphere. These models calculate air and moisture movement depending on well-established laws of physics, but struggle with processes that vary over much smaller scales such as cloud formation.
A new approach to simulating small-scale events
NeuralGCM also divides the Earth's atmosphere into cubes, and does calculations on large-scale processes such as air and moisture movement. However, it differs from traditional models by using a neural network to learn the physics of small-scale events, from existing weather data, rather than relying on simplified models. Google trained a suite of NeuralGCM models using weather data from ECMWF spanning 1979 to 2019 at various resolutions.
