Google AI on its Twitter handle announced that NeuralGCM is now available on GitHub. Global climate compartments (GCCs) are at the center of weather and climate prediction 1,2. GCMs contain a physical core which is a solver of large-scale kinematics integrated with parametrizations of the small-scale processes, for instance, cloud formation.
What is NeuralGCM?
NeuralGCM is within one standard deviation of machine-learning models for one to ten minutes forecasts and the ECMWF ensemble prediction for one to fifteen minutes forecasts. Thus, Reduced NeuralGCM can faithfully monitor climate indices for several decades, and the climate predictions with the 140-kilometre grid resolution exhibit some emergent behaviors such as the realistic frequency and paths of tropical cyclones.
How does it work?
Similar to other machine learning models of similar platforms, NeuralGCM may employ one percent, or even less, of the electricity utilized by GCMs in creating accurate short-term deterministic weather forecasts that are one to three days in the future. Google has even researched how it improves climatic conditions using AI. But, when it was giving forecasts for more than a week the number of errors was a lot considerably less than other machine learning algorithms. Indeed, ECMWF-ENS, which is a GCM widely recognized as the state-of-the-art model for weather forecasting, provided long-term forecasts similar to those of NeuralGCM.
Major features
It consists of a dynamical core derived from the discretized governing dynamical equations and a learned physics module that predicts the physics with the help of neural networks. The dynamical core predicts large-scale movement of fluids and heat under the effect of gravity and the Coriolis effect.