Beyond Prediction: Exploring Uncertainty in Climate Models in the AI Era (Core Science Keynote)

Numerical weather forecasting is quite likely undergoing a period of revolutionary change. A number of separate efforts using machine-learned emulation of historical weather reanalysis datasets such as ERA5, have shown impressive skill in forecasts of a few days. Compared with the models used to generate the reanalysis, their cost and time to solution is lower, sometimes by many orders of magnitude. Efforts are underway to learn directly from observational data streams rather than the uniform gridded reanalysis datasets generated by numerical models. It is not unlikely that we will be able to forgo the model-bootstrap step and generate global weather forecasts directly from observations using machine learning in the near future.

The extension of this approach to climate often takes the form of building foundation models trained on longer term datasets including AMIP and CMIP ensembles. The loss functions used in training often minimise trajectory deviations, such as mean square error on a timeseries of a physical variable.

We argue in this talk for a different approach to applying machine learning to the climate problem. Problems that need to be urgently addressed using these new approaches include the fate of the land and ocean carbon sinks, which might be approaching fundamental new states not seen in recent data. In this approach, the climate model is not simply a tool for prediction, but for exploration of a vast counterfactual space, much of which involves parameters, and boundary and initial conditions, outside the bounds of recent historical climate, and perhaps only accessible under theoretical idealisations. The orders-of-magnitude reduction in cost and time to solution that may be available through ML approaches is applied to the exploration of uncertainty in physics and in factors external to the climate system. We propose an experimental framework for assessing the readiness of such ML approaches to the climate problem.