Neural-Network Parameterization of Subgrid Momentum Transport Learned from a High-Resolution Simulation
Abstract
Attempts to use machine learning for developing new parameterizations have mainly focused on the effects of subgrid processes on thermodynamic and moisture variables, but parameterizations of subgrid momentum transport are also needed for operational use in weather and climate modeling. In this study we use neural networks to develop a parameterization of subgrid momentum transport that learns from coarse-grained output of a high-resolution atmospheric simulation in an idealized aquaplanet domain. We show that substantial subgrid momentum transport occurs due to convection and non-orographic gravity waves. Furthermore, we show that the zonal- and time-mean subgrid momentum tendencies in the tropical upper troposphere that are calculated directly from the coarse-grained output behave similarly to residuals from momentum budgets in reanalysis. The neural network parameterization we develop has a structure that ensures the conservation of momentum, and it has reasonable skill in predicting momentum fluxes associated with convection. However, the skill of the neural-network momentum parameterization is substantially lower compared to the skill of a neural network trained to predict moisture and energy fluxes. Predicting subgrid momentum transport might be more challenging than predicting subgrid moisture and energy transport because (a) it is difficult to predict the momentum transport by gravity waves and (b) convective momentum transport can be both negative or positive, depending on the spatial organization of clouds, while energy and moisture are always transported in the same direction during convection events. The neural-network parameterization is implemented in the same atmospheric model (that it has learned from) at coarse resolution and leads to stable simulations that do not exhibit climate drift. We find that the momentum parameterization corrects some of the biases in coarse resolution simulations relative to the high-resolution simulations although sometimes it overcorrects. Overall, our results show that neural-networks are promising for parameterizing subgrid momentum transport, and they also highlight the difficulty in parameterizing subgrid momentum transport.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2021
- Bibcode:
- 2021AGUFM.A14C..05Y