Google researchers have improved the precipitation component of NeuralGCM by training the model directly on satellite-based precipitation observations, rather than on reanalysis data. The change is aimed at giving the model a better way to represent clouds and rainfall, which are difficult for large-scale weather and climate models to resolve.
The source text says clouds can exist at scales smaller than 100 meters, far below the kilometers-scale resolution of global weather models and the tens-of-kilometers scale used in global climate models. Because clouds change quickly and involve complex physics, models cannot calculate every detail directly.
To handle those small-scale processes, models use approximations called parameterizations. NeuralGCM takes a different approach by using a neural network to learn the effects of these events directly from existing weather data.
The earlier version of NeuralGCM, like most ML weather models, was trained on reanalyses, which combine physics-based models with observations to fill gaps in data. The source says this can reproduce weaknesses in precipitation, including extremes and the daily cycle.
In the updated version, the precipitation part of NeuralGCM was trained on NASA satellite-based precipitation observations spanning 2001 to 2018. The text says NeuralGCM’s differential dynamical core infrastructure made this possible.
According to the source, previous hybrid models that combine physics and AI could only use output from high-fidelity simulations or reanalysis data. Training directly on satellite observations is described as a way to find a better, machine-learned parameterization for precipitation.
Source: research.google.
Companies can share verified announcements through Newz9’s international press release submission page.
