AI gets extreme forecasts wrong.
Keystone-SDA
Artificial intelligence (AI) is changing the face of meteorology, promising faster and more efficient forecasts. But a new study from the University of Geneva finds AI isn’t always living up to its potential.
+ Get the most important news from Switzerland in your inbox
Precisely when the weather becomes extreme and threatens the security of people and infrastructure, machine learning-based models show their Achilles’ heel. This is revealed by a new study led by the University of Geneva and published in the American journal Science Advances.
The researchers compared three of the most advanced AI systems for weather forecasting – GraphCast, Pangu-Weather and Fuxi – with the HRES physical reference model of the European Centre for Medium-Range Weather Forecasts (ECMWF). The verdict is clear: in the face of record events, the AI is systematically wrong.
According to the study, extreme cold spells are generally predicted to be less intense than they turn out to be in reality. On the contrary, heat peaks are underestimated by predicting less extreme values than in reality. The predicted temperatures are lower than the actual ones. There are also significant deviations for strong wind events.
But that is not all: the AI models not only predict the intensity of extreme events too weakly, they also predict them too rarely compared to what actually happens. “These results highlight a central challenge for the use of AI in the prediction of weather events with a large societal impact,” write the authors of the research.
More
More
Swiss scientists to use AI for improved weather and climate forecasts
In their view, the problem lies at the very heart of how AI works. Models learn from historical data: they recognise sequences and correlations that have already occurred in the past. But record events, by definition, fall outside this range of experience. What has never happened (or has happened very rarely) escapes the statistical logic of machine learning.
Physical models, by contrast, work in a radically different way. They simulate the evolution of the atmosphere based on immutable natural laws – thermodynamics, fluid mechanics – that apply regardless of whether a certain phenomenon has already been observed. “Physics does not change,” the researchers summarise. This is why traditional models are able to calculate even extreme situations never seen before.
The discovery comes at a crucial time. With climate change, extreme weather events are on the rise and with them comes the need for reliable forecasts to protect lives, infrastructure, agriculture and energy supply. “It is crucial for early warning systems that a model reliably predicts the occurrence of extreme events,” the scholars point out. If a heat wave is underestimated or a storm is recognised too late, the consequences can be severe for public health, security of supply and civil protection operations.
According to its authors, the study does not see this as a complete failure of AI in meteorology. On the contrary, the research team recognises that machine learning models offer “vast new possibilities”. The question is how to exploit its strengths without being penalised by their limitations.
“Combining both worlds in a hybrid forecasting system seems to be the best way forward,” said Roland Potthast, head of the models department at the German meteorological service DWD, when asked by the Science Media Centre, an organisation that acts as a bridge between science and the media. “In this way, the advantages of both methods could be made available to society.”
Research is already exploring practical solutions, including training approaches that give greater weight to extreme events and new types of models that can better represent uncertainties. The goal is to arrive at a system that combines the speed and efficiency of AI with the physical robustness of traditional models: a scientific and technological challenge that may benefit all citizens when faced with the next cloudburst or heat wave.
Join the debate:
Translated from Italian by AI/jdp
We select the most relevant news for an international audience and use automatic translation tools to translate them into English. A journalist then reviews the translation for clarity and accuracy before publication.
Providing you with automatically translated news gives us the time to write more in-depth articles. The news stories we select have been written and carefully fact-checked by an external editorial team from news agencies such as Bloomberg or Keystone.
If you have any questions about how we work, write to us at english@swissinfo.ch.
