Representing Model Uncertainty in Climate Ensembles (Invited Presentation)

One of Mark's most important legacies is operational seasonal climate prediction. For example, his successful predictions of the 1986/87 El Nino event kickstarted work at ECMWF on seasonal prediction, leading directly to today's European Union Copernicus Seasonal Forecast system.

Today, seasonal climate prediction is a mature science and societies around the world use seasonal forecast ensembles for a variety of applications. However, to be useful, the forecast probabilities from these ensembles must be reliable. Indeed, if forecast probabilities are not reliable, users can make inappropriate decisions. Unreliable probability forecasts can be harmful - and therefore worse than useless - to society.

One of the principal sources of forecast uncertainty is associated with the fact that our computational equations are imperfect representations of the underlying partial differential equations of the relevant laws of physics. We call this model error.

Using Navier-Stokes' theory, I shall discuss the underlying reason why model error is so pervasive in climate prediction, and on this basis review four different approaches for representing model error in forecast ensembles, each of which I have worked on in the past: multi-model ensembles, perturbed parameter ensembles, stochastic physics, and forcing singular vectors.

I will conclude that we still have a long way to go before we can say we have found a satisfactory approach to the representation of model uncertainty. To borrow from Mark's words, we should all be working on understanding better the nature of model uncertainty, and ways to represent it, given the critical importance of providing society will reliable forecasts, not just on seasonal timescales, but on longer climate change timescales and shorter weather timescales. Indeed, it is hard for me to think of anything (which spans both fundamental and applied climate science) more important for our field in the coming years - not least with the growth of AI-based forecasts.