Geophys. Res. Lett. (2017)

While considerable progress has been made in our understanding of the climate system, projections of the future remain highly uncertain. Such relatively low confidence stems, in part, from uncertainties in the parameterization schemes of Earth system models (ESMs) — approximations of unresolved small-scale processes — including, for example, cloud dynamics. Tapio Schneider and colleagues from the California Institute of Technology, USA, envision a revolution in Earth system modelling using data assimilation and machine learning to improve parameterization schemes.

With recent advances in computational tools, it is suggested that ESMs may soon be able to actively optimize parameter estimates. This can be achieved by harnessing global observations and targeted high-resolution simulations, for example in regions where parameterizations are particularly uncertain. In comparison to the traditional approach, this machine-learning-based method would systematically improve parameterizations, and in turn, reduce projection uncertainty. However, before this revolution can be achieved, advances are needed in learning algorithms, and the parameterizations themselves must be re-designed to allow for refinement using data from various sources. Nevertheless, the vast advances in computational resources may mean the re-engineering of ESMs may not be too far from realisation.