The advancements in artificial intelligence (AI) have not only revolutionized various industries but have also made their way into weather forecasting. AI weather models, such as ChatGPT, have shown promising results in predicting storm tracks and providing probabilistic forecasts. However, these models still have their limitations and challenges to overcome.
Matthew Chantry, a machine-learning coordinator at the European Centre for Medium-Range Weather Forecasts (ECMWF), acknowledges that the success of AI weather models was not initially guaranteed. Unlike other fields where AI algorithms have access to vast amounts of data, there is no comprehensive dataset for Earth’s atmosphere, especially for predicting hurricanes. Nevertheless, the accuracy of storm track predictions indicates that the algorithms have captured some fundamental aspects of atmospheric physics.
One of the drawbacks of machine-learning algorithms in weather prediction is their tendency to downplay outliers, such as extreme weather events like heatwaves or tropical storms. These models primarily focus on the most common patterns, and as a result, they may not accurately predict the intensity of these outlier events. Furthermore, the current models are not specifically designed to estimate rainfall, which unfolds at a finer resolution than the global weather data used to train them.
Predicting precipitation and extreme events are regarded as the most challenging cases for AI weather models. DeepMind, a leading AI research company, has developed a localized radar-based approach called NowCasting, which focuses on predicting rain in a more precise and localized manner. Integrating this approach with AI weather models presents its own challenges. However, with the introduction of more fine-grained data in future versions of the ECMWF data set, AI models may gain the ability to predict rainfall more accurately.
When it comes to efficiency, AI models excel in providing probabilistic forecasts. Unlike traditional physical models that can take hours to calculate multiple possible scenarios, AI models can generate multiple projections in a matter of minutes. This speed advantage makes it feasible to create large ensemble models that were previously impractical with physically based models.
However, AI models have limitations in terms of explaining their predictions. The inherent black box nature of many machine-learning systems hampers the ability to assess the uncertainty in the model itself, which is crucial in weather forecasting. Meteorologists have expressed the need for detailed explanations accompanying AI forecasts, and researchers are striving to address this request.
Despite the progress made, AI-powered meteorology is still not ready to be widely accessible. The accuracy of predictions heavily relies on reliable weather observations, which encompass data from satellites, buoys, planes, and sensors processed by organizations like NOAA and ECMWF. Intellectual property and national security concerns also play a role in limiting the availability of raw data to AI researchers, startups, and nations with limited data-gathering capacity.
While large forecasting centers continue to test AI models, meteorologists remain cautious due to the stakes involved in providing accurate forecasts that impact lives and property. Physics-based models are not expected to disappear entirely, but with ongoing improvements, it is projected that AI will play some role in official forecasts within the next couple of hurricane seasons. The potential benefits are acknowledged, and experts are optimistic about the future of AI in weather forecasting.