Modeling Atmospheric Flows with DeepMind and Graph Neural Networks: The Case of GraphCast
Gnns are networks of connected objects that can influence one another. In the case of DeepMind’s weather forecasts, each node represents a set of atmospheric conditions at a particular location, such as temperature, humidity, and pressure. There is a cloud of data distributed around the globe and at different altitudes. The goal is to know how the data will change over time, by predicting how it will interact with their neighbors.
There is still a need for standard models to provide estimates of global weather that are first used to train machine-learning models. There will be more than two to five years before people can use machine learning approaches to make decisions in the real world.
The goal of Battaglia’s team was to model the behavior of fluids with the use of graph neural networks. Given that weather prediction is at its core about modeling the flow of molecules, tapping GNNs seemed intuitive. While training these systems is heavy-duty, requiring hundreds of specialized graphics processing units, or GPUs, to crunch tremendous amounts of data, the final system is ultimately lightweight, allowing forecasts to be generated quickly with minimal computer power.
GraphCast is leading the race for the best models according to a computer scientist at University of California, Los Angeles. The model is described1 in Science on 14 November.
“In the troposphere, which is the part of the atmosphere closest to the surface that affects us all the most, GraphCast outperforms HRES on more than 99% of the 12,00 measurements that we’ve done,” says computer scientist Remi Lam at DeepMind in London. Across all levels of the atmosphere, the model outperformed HRES on 90% of weather predictions.
GraphCast predicted the state of 5 weather variables close to the Earth’s surface, such as the air temperature 2-metres above the ground, and 6 atmospheric variables, such as wind speed, further from the Earth’s surface.
It also proved useful in predicting severe weather events, such as the paths taken by tropical cyclones, and extreme heat and cold episodes, says Chantry.
Artificial Intelligence and Machine-Learning in Weather Prediction: Where We Need to Look at Next-Generation Adversarial Models
Problems with machine- learning approaches need to be fixed. Researchers can not fully understand the decision- making processes in GraphCast, because they happen in an artificial intelligence black box. “This calls into question their reliability,” she says.
Machine-learning models are still experimental, but they could be a way to boost certain types of weather prediction that conventional approaches aren’t particularly good at.
AI models also run the risk of amplifying biases in their training data and require a lot of energy for training, although they consume less energy than NWP models, says Grover.
