Sponsored by Google Cloud
Choosing Your First Generative AI Use Cases
To get started with generative AI, first focus on areas that can improve human experiences with information.
Google’s NeuralGCM is an AI model that combines machine learning and physics to deliver accurate weather forecasts
Google has unveiled NeuralGCM, a new weather and climate prediction model that combines machine learning with traditional physics-based simulators to provide accurate forecasts.
Detailed in a newly published paper in the Nature Journal, the weather prediction model can accurately predict short-term weather and long-term climate forecasts.
It uses a combination of machine learning with General Circulation Models (GCMs), which are physics-based simulators used for weather and climate predictions meteorologists have used for decades.
NeuralGCM is trained end-to-end, integrating a machine learning component directly with a GCM. The researchers say the hybrid approach enhances stability and accuracy for both short-term and long-term predictions.
NeuralGCM is a significant advancement, integrating machine learning with traditional GCMs to enhance weather and climate prediction accuracy.
The model also offers weather forecasters reduced running costs as it processes simulations in a matter of hours on a single one of Google’s custom tensor processing units (TPUs).
In their paper, the Google researchers showed NeuralGCM simulating weather conditions for 70,000 days in just 24 hours using a single TPU compared to 19 simulated days on 13,824 CPU cores. This computational efficiency can be leveraged for previously impractical tasks such as large ensemble forecasting.
“Our results show that end-to-end deep learning is compatible with tasks performed by conventional GCMs and can enhance the large-scale physical simulations that are essential for understanding and predicting the Earth system,” the researchers wrote.
Google has been developing weather prediction systems for years, leveraging them in its Search and Maps services.
Its researchers claim to have developed the world’s first machine-learning-based model to make accurate ensemble weather forecasts.
Machine learning can determine short-term forecasts but the technology struggles with longer-term predictions, providing unrealistic predictions for multi-day forecasts. NeuralGCM addresses this by delivering calibrated uncertainty estimates, essential for reliable long-term predictions.
By combining AI technology with traditional forecasting systems, the researchers were able to develop a system that can handle short and long-term predictions leveraging the best of both approaches.
The model’s key components include a differentiable dynamical core for solving large-scale dynamical equations of the atmosphere and a neural network-based learned physics module that handles processes related to cloud formations and precipitation.
Credit: Google
The team behind it claims the model can predict forecasts at a similar level to systems employed by the European Center for Medium-Range Weather Forecasts for one to 15 forecasts.
“This approach generates two to 15 day weather forecasts that are more accurate than the current gold-standard physics-based model, and reproduces temperatures over a past 40-year period more accurately than traditional atmospheric models,” said Stephan Hoyer, a senior staff software engineer at Google Research. “Although we have not yet built NeuralGCM into a full climate model, it marks a significant step towards developing more powerful and accessible climate models.”
Google isn’t keeping the model for itself. It’s making the code base publicly available through open source. NeuralGCM can be found on GitHub under an Apache license, meaning anyone can use it to create their own proprietary software.
The atmospheric dynamical core is distributed in a separate package called Dinosaur, as are the Haiku modules for defining neural network layers.
By open-sourcing NeuralGCM, Google hopes to foster further innovation and collaboration within the meteorological and climate science communities.
However, the team acknowledged on the API page that modifying and fine-tuning the related models is “trickier than it needs to be,” with plans to refactor the model’s code for improved usability.
You May Also Like