AI is Revolutionising Weather Forecasting

AI-based weather forecasting models, like AIFS and WeatherMesh, are outperforming traditional physics-based models in certain scenarios.,The European Centre for Medium-Range Weather Forecasts' ERA5 dataset, with its rich atmospheric data, is a key resource for training AI weather models.,WindBorne Systems, an innovative start-up, is enhancing weather forecasting by launching small, long-lasting weather balloons to gather global atmospheric data.

A New Era of Weather Forecasting: The AI Advantage

The weather forecasting community is undergoing a significant transformation, thanks to the advent of artificial intelligence (AI). This revolutionary technology is paving the way for a new method of weather forecasting that can operate on a simple desktop computer.

Traditional AI systems rely heavily on data to function effectively. Large language models, such as ChatGPT, consume vast amounts of data to improve their responses to user queries. However, the availability of high-quality data is limited, even on the internet. To overcome this, operators of AI models are exploring the use of synthetic data and other untapped data sources. For more on the challenges of data scarcity, read about Running Out of Data: The Strange Problem Behind AI's Next Bottleneck.

One such promising data source is weather forecasting. The European Centre for Medium-Range Weather Forecasts (ECMWF), a leading organisation in numerical weather prediction, maintains a comprehensive dataset called ERA5. This dataset contains atmospheric, land, and oceanic weather data from 1940, with particularly rich data from the last 50 years due to global satellite coverage. Although not initially intended for AI applications, ERA5 has proven to be incredibly valuable for training AI weather models.

AI weather models have rapidly progressed since computer scientists began utilising ERA5 in 2022. In some instances, these models have even surpassed the accuracy of traditional physics-based global weather models, which have taken decades to develop and require powerful supercomputers to run.

Matthew Chantry, who heads AI forecasting efforts at ECMWF, affirms that "machine learning is a significant part of the future of weather forecasting."

WindBorne Systems: Enhancing Weather Forecasting with Innovative Technology

John Dean and Kai Marshland, two Stanford University undergraduates, co-founded WindBorne Systems with the aim of tackling the issue of weather uncertainty. The company's premise is simple: to gather quality weather data from the 85% of the Earth's atmosphere that lacks it.

Traditional weather balloons, which provide valuable atmospheric data, are cumbersome and only function for a few hours. The National Weather Service in the United States launches them twice daily from around 100 locations. To overcome this limitation, Dean and Marshland developed smaller, lighter balloons that can persist in the atmosphere for up to 40 days. By launching hundreds of these balloons each day, WindBorne Systems has amassed a wealth of atmospheric data from around the globe.

To incorporate this balloon data into forecast models, WindBorne Systems began developing its own AI-based weather model, WeatherMesh, about a year ago. WeatherMesh has since outperformed traditional physics-based models in tasks such as hurricane forecasting. The company now offers both balloon data and the impressively accurate WeatherMesh model to its customers.

The Origins and Future of AI Weather Forecasting

Academic work on using deep learning techniques for weather forecasting began around six years ago. This form of machine learning, inspired by biological brains, uses neural networks to identify and classify information, recognise patterns, and explore possibilities. For more insights into the future of AI, consider Adrian's Angle: AI in 2024 - Key Lessons and Bold Predictions for 2025.

Initially, computer scientists were sceptical about the effectiveness of this approach, as it differed greatly from the established science of weather forecasting. However, in 2022, promising results from the use of graph neural networks and the Chinese-based Huawei's Pangu-Weather model demonstrated that AI-based models could, in certain scenarios, outperform the ECMWF's physics-based model, which is considered the best in the world. A detailed comparison of AI weather models and traditional methods can be found in a recent article from the American Geophysical Union.

These findings sparked a wave of interest in the development of AI weather models. Chantry and his colleagues at ECMWF began exploring the possibilities in early 2023, and by the end of that year, the AIFS (Artificial Intelligence/Integrated Forecasting System) model was already producing encouraging results. In the spring of 2024, ECMWF started publishing real-time AIFS forecasts, which have since become an increasingly useful tool for meteorologists.

While physics-based weather models are still widely used and trusted, the future of weather forecasting is likely to involve a combination of both AI and traditional methods. Chantry and his team are currently working on techniques to allow AI models to ingest current observations, potentially enabling them to perform both data assimilation and forecasting. This, he says, is a more challenging problem than training AI models, but one that could revolutionise the field of weather forecasting. This shift highlights a broader trend in AI's Secret Revolution: Trends You Can't Miss.

Comment and Share:

What are your thoughts on the growing role of AI and AGI in weather forecasting? Do you believe these technologies have the potential to significantly improve the accuracy and efficiency of weather predictions? We'd love to hear your opinions in the comments below.

Don't forget to Subscribe to our newsletter for updates on the latest AI and AGI developments in Asia and beyond.