Oxford Professor Philip Stier: AI Models Face Fundamental Difficulties in Predicting Out-of-Distribution Extreme Events
Philip Stier, Professor of Atmospheric Physics at the University of Oxford and Director of the Intelligent Earth Centre, explains on a Man Group podcast why AI climate models struggle to predict extreme events that have never appeared in their training data, and how biases in reanalysis data combined with large-sample ensembles can create a false sense of precision.
At the University of Oxford, Professor of Atmospheric Physics and Director of the Intelligent Earth Centre, Philip Stier, discussed with Jason Mitchell, Chief Investment Officer for Responsible Investment at Man Group, how artificial intelligence is transforming climate modeling during an episode of the firm's "A Sustainable Future" podcast. Stier’s core assessment is that purely data-driven AI models face fundamental difficulties when predicting out-of-distribution extreme events, a limitation that cannot be overcome simply by increasing the volume of training data or running more simulations.
Stier defines the challenge of predicting climate change as an "out-of-distribution problem": we are required to forecast states that have never occurred in the training dataset. Traditional physical models encode physical laws through systems of equations; these equations remain valid even under unprecedented conditions. In contrast, AI models learn patterns from historical observations. When future states exceed the coverage of the training data, the model lacks a basis for extrapolation. He cites the current El Niño event as an example: because observational records are too short, such extreme ENSO (El Niño-Southern Oscillation) events may not be present in the training data, making pure AI-driven climate predictions unreliable in such scenarios.
A deeper issue lies in the biases inherent in the training data itself. Stier points out that most AI weather models currently rely on reanalysis data for training—a combination of historical observations and outputs from traditional physical models. This means AI inherits not only the flaws of observational data (such as insufficient coverage in developing countries and limited funding for public observation systems) but also the parametrization biases of the physical models themselves. For instance, reanalysis data does not idealize the diurnal cycle of precipitation, which is related to the parametrization of convective clouds; AI models will learn this same bias. While acceptable for short-term weather forecasting provided predictive skill improves, this inherited bias can distort judgments on key processes like cloud feedback in climate prediction.