A.I. is Quietly Powering a Revolution in Weather Prediction
In February, the European Centre for Medium-Range Weather Forecasts (ECMWF), a global leader in forecasting weather conditions weeks in advance, launched the world’s first fully operational weather forecast system powered by artificial intelligence (A.I.). This milestone in meteorology marks a significant turning point in how weather predictions are generated and utilized globally.
The A.I. forecasts are remarkably more efficient—being produced faster, easier, and with much less computational energy than traditional methods, requiring 1,000 times less energy. Moreover, these machine learning-driven forecasts have shown greater precision in accuracy. “Currently, the machine learning model is outperforming traditional models,” says Peter Dueben, a model developer at ECMWF in Bonn, largely thanks to the center’s Artificial Intelligence Forecasting System (AIFS). Some weather phenomena predictions have already seen a 20% improvement over previous state-of-the-art models.
Andrew Charlton-Perez, a meteorologist from the University of Reading and head of the school’s computational sciences department, anticipates an increase in operational A.I. forecasts from both national agencies and tech giants like Google. “The field is advancing at an incredible pace,” he comments, suggesting the transformative power A.I. holds for meteorology.
Artificial intelligence is driving a revolutionary shift in how weather forecasts are developed, enabling processes that once needed large teams and supercomputers to now be feasible on a laptop. This innovation is particularly beneficial for under-resourced countries, empowering them to create their own forecasts without needing high-end technology. Researchers can also explore vast datasets to understand the underlying reasons for A.I.’s efficacy, potentially uncovering unknown physics in meteorology not yet accounted for in traditional models.
However, the deployment of A.I. in weather prediction isn’t without its challenges. The “black box” nature of A.I.—where decision-making processes are not always transparent—poses trust issues among meteorologists. The increasing unpredictability of weather patterns due to climate change presents additional risks, as these systems may falter under conditions outside their training data.
Presently, meteorologists rely on both A.I. and traditional forecasts, with the former benefiting from the data produced by the latter’s slower supercomputers. A critical question looms: will A.I. evolve to the point where it becomes the sole requirement for accurate weather forecasting?
Despite the skepticism some hold towards the reliability of weather forecasts, they have improved significantly over the past half-century. As Dueben mentions, forecasts now gain roughly a day’s worth of accuracy each decade; today’s six-day forecast is as precise as what a five-day forecast was ten years prior.
Traditionally, weather prediction involved analyzing weather systems and historical patterns. By the early 20th century, a shift toward physics-based models emerged, aiming to predict atmospheric behaviors through fluid dynamics and thermodynamics. The significant advancements in computing in the 1960s saw the advent of numerical weather prediction. The establishment of the ECMWF in 1975 pooled European resources to generate forecasts using these advanced techniques, a task reliant on extremely powerful computers at the time.
As observational data has increased—now exceeding 200 billion data points daily from various sources—so have computer capabilities, creating finer, higher resolution models. Advanced national services refine global models for more detailed local forecasts, even achieving resolutions down to 100 meters in some regional models experimented with by the U.K.’s Met Office.
This incremental progress is referred to as a “quiet revolution” in the sciences, and now A.I. is marking the onset of what Charlton-Perez describes as the “second revolution.” Machine learning technologies excel at recognizing complex patterns in large datasets, reminiscent of historic forecasting strategies rooted in past weather patterns.
Recent years have seen notable tech companies and academic researchers develop A.I.-based weather systems. Google Deepmind’s GraphCast, CalTech’s FourCast, and Huawei’s Pangu-Weather systems are examples. While not yet operational for routine forecasts, these systems often match the accuracy of traditional numerical predictions with significantly less computational demand.
The ECMWF has tested A.I.-based systems extensively, with satisfactory results even during rare extreme events. In example, during Storm Ciarán in 2023, A.I. achieved impressive forecasting accuracy comparable to traditional systems, though challenges remain in capturing finer ground-level variables precisely.
The ongoing evolution has yielded even more advanced systems, like Google’s GenCast, which provides probabilistically varied forecasts to predict event likelihoods. This progress aligns closely with ECMWF’s efforts towards their own ensemble forecasts.
Despite the promise A.I. holds, it is vital to continue leveraging traditional physics-based models to support and refine A.I. efforts, evident in the reliance on the ECMWF’s ERA5 dataset. However, ground-breaking A.I. models like Aardvark are emerging, leveraging solely observational data to create forecasts, hinting at a potential shift away from physics-based models.
Such developments are complemented by ongoing research into explainable A.I. (X.A.I.), a promising field proposed to demystify A.I. processes. These sophisticated “brain surgeries” on A.I. models may unlock new understanding into meteorological phenomena and usher in improved forecasting models.
A.I. provides only one aspect of the future of weather prediction. While it may simplify forecast production, making it accessible for all nations, it isn’t expected to replace the nuances of physical models or human expertise anytime soon. As Ramsdale highlights, the human element is crucial in interpreting data into actionable advice.
