Deep Learning to Represent Subgrid Processes in Climate Models
Most earth system models still rely on imperfect parameterizations of unresolved cloud processes. SuperParameterization (SP) – embedding thousands of limited-domain cloud-resolving arrays in a global climate model -- is a promising alternative but its computational cost has been prohibitive. Here researchers from UC Irvine, Ludwig Maximilians University, and Columbia, show that SP calculations can be successfully emulated with a simple deep neural network, which even displays some potential for out-of-sample generalizability.
Limited ability to access computing power remains the main barrier to including satisfying representations of cloud physics in modern climate models. Finding ways to emulate prognostic cloud-resolving calculations with compact matrix calculations, i.e. through deep learning, is a tantalizing idea in this context. The current study published in Proceedings of the National Academy of Science proves that a deep learning algorithm trained on SP is capable of interacting appropriately when allowed to feed back with a fully prognostic global dynamical core, in an idealized aquaplanet setting. Successes include the emergence of realistic equatorial wave spectra, rainfall extremes and correct generalization to a zonally asymmetric surface temperature perturbation beyond the training dataset’s boundaries. Limitations of generalizability to more extreme forms of surface warming are also discovered, as well as numerical stability issues, which highlight the importance of improving our understanding of the trade-offs of neural network informed subgrid parameterization in the coming decade.
Replacing cloud superparameterization with a deep neural network trained on its essence, we show it is possible to create a global climate model that reproduces the essence of explicit convection calculations at a small fraction of their actual cost. Our analysis affirms the potential of deep-learning emulation for subgrid parameterization.