Artificial Intelligence (AI) firm Google DeepMind has made a significant leap in weather forecasting by introducing its machine-learning model, GraphCast. This innovative model has outperformed traditional tools and surpassed other AI approaches in the complex task of weather prediction.
GraphCast, hailed as a leader among AI models, can run on a desktop computer, providing predictions that are more accurate and faster than those generated by conventional models. The model’s capabilities were detailed in a Science article on November 14, drawing attention to its potential to revolutionise the landscape of weather forecasting.
Also Read: 5 Groundbreaking Benefits of Using Shield AI in Defence and Security
The Complexity of Weather Prediction
Weather forecasting has long been challenging and resource-intensive, traditionally relying on Numerical Weather Prediction (NWP). NWP involves running physical models on supercomputers and processing data from various sources such as buoys, satellites, and weather stations worldwide.
While these physical models accurately map the movement of heat, air, and water vapour in the atmosphere, they come with high financial and energy costs.
In response to these challenges, several technology companies, including DeepMind, NVIDIA, and Huawei, alongside start-ups like Atmo, have turned to machine-learning models. These models leverage past and current weather data to predict the future state of global weather rapidly. Among them, Huawei’s Pangu-Weather model has emerged as a strong rival to the gold-standard NWP system at the European Centre for Medium-Range Weather Forecasts (ECMWF), providing world-leading weather predictions up to 15 days in advance.
Also Read: Answers AI Solves All Problem with 3 Shocking Benefits
Machine Learning: A Revolution in Weather Forecasting
Machine learning has spurred a revolution in weather forecasting, offering a faster alternative to conventional NWP models. Matthew Chantry, a machine-learning coordinator at ECMWF, notes that AI models run 1,000 –10,000 times faster than their traditional counterparts. According to Jacob Radford, a data-visualisation researcher at the Cooperative Institute for Research in the Atmosphere in Colorado, this increased speed allows scientists more time to interpret and communicate predictions.
Presenting GraphCast: our state-of-the-art AI model delivering 10-day weather forecasts with unprecedented accuracy in under one minute. 🌦️
— Google DeepMind (@GoogleDeepMind) November 14, 2023
It can even help predict the potential paths of cyclones further into the future.
Here's how it works. 🧵 https://t.co/ygughpkdeP pic.twitter.com/0Y6DyBXDow
GraphCast, the latest addition to machine-learning models, has demonstrated its superiority over conventional and AI-based approaches in global weather forecasting tasks. Researchers trained the model using estimates of past global weather from 1979 to 2017, allowing GraphCast to learn the intricate links between weather variables such as air pressure, wind, temperature, and humidity.
Related: Machine Learning (ML) vs. AI: 3 Great Talking Points
GraphCast: A Technological Marvel
GraphCast operates on machine learning and Graph Neural Networks (GNNs), proving to be a particularly valuable architecture for processing spatially structured data. The model produces forecasts at an impressively high resolution of 0.25 degrees longitude/latitude (28km x 28km at the equator), covering more than a million grid points across the Earth’s surface. GraphCast predicts five Earth-surface variables and six atmospheric variables at each grid point at each of the 37 altitude levels.
While the training process for GraphCast was computationally intensive, the resulting forecasting model is highly efficient. Making 10-day forecasts with GraphCast takes less than a minute on a single Google TPU v4 machine. A 10-day forecast using a conventional approach, such as HRES, could take hours of computation on a supercomputer with hundreds of machines.
Computer scientist Rémi Lam at DeepMind in London said, “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 12,000 measurements that we’ve done.” Across all levels of the atmosphere, the model outperformed HRES on 90% of weather predictions.
Also Read: AI-Powered Virtual Hosts Disrupt China’s E-commerce Market
Advantages and Applications
GraphCast has proven effective not only in predicting standard weather variables but also in forecasting severe weather events. This includes predicting the paths taken by tropical cyclones and forecasting extreme heat and cold episodes. In a comparative study with Huawei’s Pangu-Weather, GraphCast outperformed 99% of weather predictions from a previous Huawei study.
The application of GraphCast extends beyond its predictive accuracy. It excels in predicting severe weather events, making it a valuable tool in scenarios such as tropical cyclone paths and extreme temperature episodes. Matthew Chantry emphasises that it remains an experimental model despite GraphCast’s superior performance in certain metrics. Future assessments using different metrics may yield varying results.
However, integrating machine-learning models like GraphCast does not signify the complete replacement of conventional approaches. Chantry notes that while machine-learning models may excel in specific types of weather prediction, such as short-term rainfall forecasts, they cannot fully replace standard physical models. These traditional models are still essential for providing initial estimates of global weather, which are crucial for training machine-learning models.
Also Read: 5 Best Benefits of AI Mini Drone in Life-Saving Scenarios
Future Challenges and Considerations
The promising capabilities of GraphCast do not come without challenges. Researchers face the inherent complexity of machine-learning models, where decision-making processes occur within a ‘Black box‘. Unlike NWP models, understanding how AIs like GraphCast arrive at their predictions is challenging, raising questions about their reliability.
Chantry anticipates that it may take another two to five years before machine-learning approaches can be effectively used for real-world decision-making. During this time, researchers aim to address the current limitations and enhance their understanding of machine-learning models. The ongoing dialogue surrounding the reliability and interpretability of AI models underscores the need for careful consideration as these technologies continue to evolve.
Also Read: Here’s Why the Global Artificial Intelligence (AI) Market Is Fluctuating
Final Say
The introduction of GraphCast by Google DeepMind marks a significant milestone in integrating artificial intelligence into weather forecasting. With its ability to provide accurate predictions up to ten days in advance in less than a minute, GraphCast showcases the potential of machine-learning models in transforming our approach to weather prediction. As the complexities of weather forecasting persist, GraphCast stands as a testament to the evolving landscape of technology and its role in enhancing our understanding of the natural world.
You Might Also Want to Check Out: 5 AI Movies That Show How AI Might Destroy the World