Predictability of midlatitude weather: Butterflies, error growth and AI models

Professor George Craig (LMU, Munich) Cambridge Fluids Network - fluids-related seminars 20 September 2024 3:30pm MR2, Centre for Mathematical Sciences, Wilberforce Road, Cambridge Numerical weather forecasts have been improving steadily for 50 years, and this improvement has accelerated in the last two years through the application of machine learning to weather prediction. But predictability is believed to be intrinsically limited by the chaotic dynamics of the underlying equations of motion - the butterfly effect. This talk will briefly review the Lorenz theories of predictability, and present some recent research to determine more precisely how they apply to the atmosphere. With this knowledge we evaluate the performance of four recent AI weather models at the limit of predictability.