For most of human history, weather has been the great unknown that caused humans to tremble and civilisations to teeter. Whether there would be food or famine that year depended on rainfall or an untimely frost. That is why weather forecasting has been a goal for millennia; today, the problem is essentially solved, but a new revolution is now underway thanks to technology that is already present in so many fields: artificial intelligence (AI).
In the days before modern science, meteorology was steeped in superstition, religion and folklore. Answers were sought in astrology; rogations, including processions, beseeched divine intervention to bring rain, and popular myths, lacking any credible foundation or reliability—like Spain’s cabañuelas—were fervently believed. Nevertheless, even in ancient times there were efforts to systematically understand meteorological phenomena. More than six centuries before our time, the Babylonians scrutinised cloud patterns and optical phenomena such as halos in an attempt to predict the weather. Similar endeavours took place in the East, and in classical Greece, the philosopher Aristotle penned his treatise Meteorologica around 340 BC, in which he blended acute observations with gross errors.
Patterns were also the basis of early modern forecasting methods. In the 19th century, detailed records, aided by measuring instruments, were used to predict the weather when conditions similar to past ones appeared. The invention of the telegraph made it possible to send forecasts over long distances to assist navigators. The first forecasts are attributed to British naval officers Francis Beaufort and Robert FitzRoy—captain of HMS Beagle during Charles Darwin’s renowned voyage in the 1830s. FitzRoy set up what was to become the UK Met Office in 1854. The Times published its first weather forecast in 1861.
Numerical models for forecasting
In the 20th century, Norwegian physicist Vilhelm Bjerknes and English mathematician Lewis Fry Richardson laid the foundations for today’s forecasting methods: numerical models using equations that describe the workings of the atmosphere. At the time, however, knowledge was still limited. Richardson envisioned a huge team of technicians performing the necessary calculations in a chain, an impractical concept that was supplanted by computing. In 1950, the ENIAC, the first general-purpose programmable electronic computer, produced the first computerised numerical weather forecast, a 24-hour forecast that took the machine nearly as long to complete. By 1955, forecasting using these models had begun to evolve into a practical tool.
These predictive models, refined and improved over time, now serve as the cornerstone of modern forecasting. According to atmospheric scientists Russ Schumacher and Aaron Hill of Colorado State University, today’s forecasts can predict heavy rainfall two days in advance with the same precision that was achievable only for the same day back in the mid-1990s. And over the past three decades, errors in predicting hurricane tracks have been cut in half. But there is still room for improvement when it comes to predicting extreme events, and the chaos factor makes forecasting beyond 10 days difficult. In addition, forecasting requires an immense amount of data from ground stations and satellites for each calculation, necessitating the division of the atmosphere into a vast grid for equation solving. The computations, in turn, require hours of processing by supercomputers, limiting forecasts to about four times a day.
New algorithms
In contrast to these traditional systems, a new revolution is emerging through machine learning. Interestingly, AI algorithms are returning to Beaufort and FitzRoy’s concept: comparing patterns from the past, but with a capacity that is unattainable for humans. In 2023, Google DeepMind introduced its GraphCast system and Huawei unveiled Pangu-Weather. These AI algorithms require complex learning; GraphCast, for example, was trained on 40 years of historical data on 32 computers over four weeks.
But once trained, the algorithm can run on a desktop computer. According to DeepMind scientists, the system “predicts hundreds of weather variables for the next 10 days at 0.25° resolution globally in under 1 minute,” outperforming numerical models in 90% of cases, even for extreme events. All this at an energy cost that is 1,000 times cheaper than using numerical models on supercomputers. The European Centre for Medium-Range Weather Forecasts, whose numerical model is considered the best in the world, has started to introduce AI forecasting on an experimental basis. Google also has a system, MetNet-3, for high-resolution 24-hour forecasts.
According to experts, AI will not displace numerical models in the short term. Schumacher and Hill emphasise that forecast systems solely relying on historical data and not bound by the equations of atmospheric dynamics may yield unrealistic results—similar to AI-generated imagery; after all, as the journal Science points out, a trained AI is a black box. “Finding the right balance between automated tools and the knowledge of expert human forecasters has long been a challenge in meteorology,” say the two researchers.
But the progress of these models will also benefit the broader field of climate and climate change. According to Science, these AI systems will drive a new generation of high-resolution climate models running on new exascale supercomputers. This will generate enough climate data to then train AIs capable of predicting climate evolution with significantly enhanced long-term accuracy.
Comments on this publication