NeuralGCM update: new models, new license, new datasets

[NeuralGCM]

Happy holidays from Google’s NeuralGCM team!

We have been hard at work this year on improving NeuralGCM, and have some exciting updates to share:

☔ New models that predict precipitation

Our new paper Neural general circulation models optimized to predict satellite-based precipitation observations is now available on the arXiv. This paper presents new NeuralGCM models at 2.8° resolution that predict precipitation. The new models have less bias, more realistic extremes, a more accurate daily cycle and better weather forecasts of precipitation than traditional atmospheric models.

🔄 New license for model checkpoints

Pre-trained NeuralGCM checkpoints (including our new precipitation models) are now available for download under the CC BY-SA 4.0 license, which removes the previous “Non-Commercial” restriction. We want everyone to be able to use NeuralGCM!

🌎 New dataset release of climate simulations

NeuralGCM atmosphere-only climate simulations, from both the original NeuralGCM paper and this new paper on predicting precipitation, are now available for download on Google Cloud Storage.

What’s next for NeuralGCM?

We are currently focused on developing improved finer-resolution stochastic models that predict important surface observations (such as precipitation) and that have suitable stability for seasonal forecasts and climate simulation studies. We are also in the midst of refactoring our model, which should make NeuralGCM easier to customize and extend.

For technical questions on using NeuralGCM, please file an issue on GitHub. You can also reach out to the NeuralGCM team directly at neur...@google.com.

Sincerely,

Stephan Hoyer on behalf of the NeuralGCM team