High resolution (cloud resolving) simulation is known to fix many of the problems associated with imperfect parameterizations of subgrid convection in climate models. But this is still computationally impractical for climate simulation. In this new GRL paper, we attempt to extract the essence of high-resolution physics using deep learning applied to output from short, global, superparameterized simulations. A fully connected neural network is found to skillfully predict moistening and heating from explicit embedded convection arrays when trained on data from short (3-12 month) integrations comprising 35M-140M training samples. The implication is that a new machine-learning era for improved subgrid parameterization in climate modeling may soon be possible.
Gentine, P., M. S. Pritchard, S. Rasp, G. Reinaudi and G. Yacalis (G), 2018. Could machine learning break the convection parameterization deadlock?, Geophysical Research Letters, 45. https://doi.org/10.1029/2018GL078202