Assessing the potential of deep neural networks for emulating cloud superparameterization in climate models under real geography boundary conditions

Abstract:

There is growing recognition that modern machine learning networks trained on high-resolution simulations with explicit convection may sidestep historical problems with parameterized convection in climate models. While trade-offs have been explored thoroughly in aquaplanets there has been no testing to date in realistic settings i.e. including continents. We explore the potential of feed-forward deep neural networks (DNNs) for emulating cloud superparameterization in an operational climate modeling setting, using training data generated from a modern version of the Super Parameterized Community Atmospheric Model (SPCAM5). Unlike in previous idealized aqua-planet testbeds, we find our DNN hyperparameters must be semi-automatically tuned in hundreds of trials to identify networks that successfully learn the patterns within the data, and that higher data volume ~9 years (3.9 Terabytes prior to selective sampling) are required for accurate training. Even faced with the additional complexity introduced by seasonality and the diurnal cycles of continents, DNNs produce satisfying hold-out validation error rates for convective heating and moistening, competitive with past aquaplanet studies. The DNN emulator, given correct large-scale inputs, correctly fits geographically complex convective response signals including the diurnal cycle of convection and its land-sea contrast. However, representation of the tropical boundary layer remains imperfect at model time step interval (15 minutes) consistent with an overall struggle of largely deterministic DNNs to capture the most rapidly varying stochastic boundary layer signals. The results reinforce the promising potential of DNNs to represent the deterministic component of sub-grid convection in next-generation climate models, even in situations of realistic complexity.