The AI revolution in climate science

We have just witnessed the start of a paradigm shift in earth science. A paper published in Nature in July showed that a neural network (artificial intelligence) predicted weather better than the European Centre for Medium-Range Weather Forecasts, which has the world’s most advanced forecasting system. Then, in November, Google’s DeepMind announced that its weather-forecasting AI had produced even stronger predictions.

The traditional approach to weather forecasting is to use observations taken at a point in time as initial conditions for equations based on physical principles. By contrast, an AI will ingest data collected over long periods of time and then “learn” the dynamics that traditional equations must describe explicitly. Both the traditional and the AI-based method rely on supercomputers, but the AI has no need for formally developed theories.

Weather forecasting determines when and where planes fly, which routes ships take, and helps manage all manner of civilian and military risks that come with a variable environment. It matters. While these are still relatively early days for AI applications in this field, and much still needs to be worked out, as in other sectors, AI-driven forecasting may displace skilled labour, since neural networks don’t require knowledge of dynamical meteorology (the authors of the Nature paper are engineers with no such background). But the implications hardly stop there.

Writing about the problem of statistical forecasting in the 1950s, Norbert Wiener, the father of cybernetics, pointed out that if we already know the history of a system that exhibits certain properties, adding knowledge of the equations that govern its dynamics will not necessarily improve our predictions. Wiener was making a largely theoretical point, because limitations in observations, data, computing power, and other factors allowed for nothing else at the time. But now, his argument strikes at the heart of the matter, capturing the broader implications of recent advances in AI.

Just in the last few years, we have vastly increased our observational data of Earth. Between 1993 and 2003, just 25 Earth-observing satellites were launched into space; but between 2014 and 2022, the number shot up to 997, bringing the total fleet of Earth-observing and other satellites currently in orbit to some 7,560. With a vast space infrastructure that beams down data on just about anything – from plant growth, water vapour, and infrastructure installations, to infrared radiation, canopy height, and measurements of the state of the atmosphere – we have entered a golden age of Earth observation.

This growing archive of data describes almost everything that both we and nature do on Earth. When combined with new AI models and our ever-expanding computational infrastructure, it could upend our understanding of the planet and our role on it.

Consider climate change. For the last 40 years, humanity’s response to the climate crisis has been guided by the Intergovernmental Panel on Climate Change, a scientific body that is compartmentalized by discipline: the physical sciences use large Earth system models that have much in common with those used in weather forecasting, while, separately, economists and geographers quantify the impact and focus on the role of adaptation and mitigation policies on our societies.

This division of labour – reflected in the IPCC’s tripartite working groups – corresponds with a division in methodologies. While physics-based Earth system models follow from first-principles equations, economists and impact modellers use a portfolio of empirical methods and irreducible theories.

Giulio Boccaletti, Scientific Director of the Euro-Mediterranean Center on Climate Change, is the author of Water: A Biography (Pantheon Books, 2021) and Siccità (Mondadori, 2023).