Published on November 17, 2023, 2:14 pm

Data from a long-term simulation called ERA5 has been used to develop a new AI program called GraphCast, which is capable of predicting weather conditions more accurately than traditional models. GraphCast, developed by Google’s DeepMind, utilizes the vast amount of climate data collected by climatologists over several decades. By analyzing this data using the power of artificial intelligence, GraphCast can generate predictions about temperature, wind speed, air pressure, and other variables on an hourly basis.

The breakthrough achieved by DeepMind scientists demonstrates the potential of AI in revolutionizing weather forecasting. Although GraphCast is not meant to replace traditional forecasting models, it serves as a valuable complement to existing methods. The success of GraphCast is credited to the precise efforts made by human climate scientists in creating algorithms that reconstructed historical weather data with great accuracy. Without this foundation, GraphCast would not have been possible.

To test the effectiveness of their program, DeepMind researchers compared its predictions to those generated by HRES, a widely used model developed by ECMWF (European Centre for Medium-Range Weather Forecasts). HRES typically forecasts weather conditions worldwide for the next 10 days within an area measuring around 10 square kilometers. However, it relies on mathematical models created by expert researchers and requires multi-million-dollar supercomputers for processing.

GraphCast takes a different approach by representing weather data as individual points linked to neighboring areas through a graph structure. This graph neural network is trained to understand the relationships between these points and how they evolve over time. The research team trained GraphCast using 39 years’ worth of ERA5 data and then evaluated its predictive abilities against HRES.

The results were impressive: GraphCast outperformed HRES in approximately 90% of prediction tasks. It excelled in predicting extreme hot and cold phenomena as well as changes in the stratosphere but struggled when making predictions beyond a 10-day window due to increased uncertainty. Overcoming this limitation and generating probabilistic forecasts that explicitly account for uncertainty is the next crucial step for DeepMind.

While GraphCast is currently limited to historical data and not used in real-time weather prediction, DeepMind envisions its broader application beyond just weather forecasting. The program has the potential to be applied to other relevant fields, including climate and ecology, energy, agriculture, human and biological activity monitoring, and complex dynamical systems. By training simulators like GraphCast with rich real-world data, machine learning can significantly contribute to advancing physical sciences.

In summary, GraphCast represents a significant milestone in weather forecasting powered by generative AI. By leveraging decades of climate data and implementing a graph neural network approach, DeepMind’s program demonstrates superior accuracy in predicting various weather conditions. Furthermore, this breakthrough paves the way for broader applications of AI in understanding and predicting complex systems across different domains.


Comments are closed.