DeepMind’s GraphCast AI predicts weather more accurately than traditional models
Google’s DeepMind has developed a new AI model that can predict weather conditions more accurately than traditional models running on supercomputers.
The multi-decade simulation called ERA5, kept record of measurements such as wind speed, air pressure, and temperature, which are used by the GraphCast graph network to predict future measurements at different points on earth.
The AI model, GraphCast neural net, outperforms the traditional HRES model to predict temperature, air pressure, and wind speed using data from the ERA5 weather records. Although it’s not actively predicting the weather in production, it scored high on a controlled experiment with data that simulates the weather.
The GraphCast model takes a month to train using 32 TPU chips and it displays blurry weather predictions when the predictions go beyond the 10 days.
Google’s DeepMind has plans to use GraphCast and other AI models to predict other natural phenomena, such as climate and ecology, energy, agriculture, and human and biological activity, to advance the role of machine learning in the physical sciences.
DeepMind’s AI is surpassing traditional weather forecasting technology
