Google has unveiled a new weather prediction model dubbed NeuralGCM, which combines machine learning (ML) with conventional techniques to better simulate the Earth’s atmosphere.
NeuralGCM has been developed in partnership with the European Centre for Medium-Range Weather Forecasts (ECMWF). The model brings together established physics-based modelling and ML for enhanced simulation accuracy and efficiency.
According to Google Research senior staff software engineer Stephan Hoyer, NeuralGCM is yet to be integrated into a full climate model. However, it represents a major step towards developing more powerful and accessible climate models.
The approach adopted by NeuralGCM produces weather forecasts for two to 15 days that surpass the accuracy of the existing gold-standard physics-based model. The model also reproduces temperatures over a 40-year historical period more accurately than conventional atmospheric models.
NeuralGCM segments the atmosphere of the Earth into cubes and conducts calculations on the physics of large-scale processes such as air and moisture movement.
Unlike conventional models that rely on parameterisations formulated by scientists to simulate small-scale aspects like cloud formation, NeuralGCM employs a neural network to learn the physics of these events from existing weather data.
Google Research senior staff software engineer Stephan Hoyer said: “NeuralGCM currently models just Earth’s atmosphere. We hope to eventually include other aspects of Earth’s climate system, such as oceans and the carbon cycle, into the model.
“By doing so, we’ll allow NeuralGCM to make predictions on longer timescales, going beyond predicting weather over days and weeks to making forecasts on climate timescales.
“NeuralGCM presents a new approach to building climate models that could be faster, less computationally costly, and more accurate than existing models.”
Google has trained a set of NeuralGCM models by leveraging ECMWF weather data from 1979 to 2019 at resolutions of 0.7°, 1.4°, and 2.8°.
While the models were trained on weather forecasts, NeuralGCM was designed to serve as a general-purpose atmospheric model, said Hoyer.
