[AAS01-11] Parameterization of Subgrid Momentum Transport Using a Global Cloud Resolving Model and AI
★Invited Papers
Keywords:global cloud resolving model, Artificial Intelligence, parameterization
The effects of subgrid momentum transport were parameterized using the simulation results of the high-resolution global nonhydrostatic model NICAM with the aid of the so-called Artificial Intelligence (AI) technology. Using the 30 days simulation of the 14km-mesh meso-scale resolving simulation data by NICAM, the subgrid transports are evaluated in the 5.625 deg × 5.625 deg grid-mesh in the tropics between 22.5N and 22.5S. The source term of grid-scale kinetic energy E and the shear production term by the subgrid vertical momentum transport S are evaluated in the tropical domain between 30N and 30S.
A convolutional neural network (CNN) is used as machine learning. Vertical profiles of E and S are estimated by inputting horizontal velocity ( u,v) and ground surface temperature T of the grid considered with surrounding points (total 9 points). To simplify output data, we selected four layers at the altitudes 2, 4, 6, and 8 km and only signs of E and S are considered. Thus, the output profile is classified into the 16 patterns with the binaries of the four layers.
The results show that the accuracy of prediction of the vertical profile pattern of E and S by the learned AI for the test data was about 20 ~ 25%. It was found that the AI parameterization is possible to roughly reproduce the geographical distribution of E and S in the one-month average as a climatological sense.
A convolutional neural network (CNN) is used as machine learning. Vertical profiles of E and S are estimated by inputting horizontal velocity ( u,v) and ground surface temperature T of the grid considered with surrounding points (total 9 points). To simplify output data, we selected four layers at the altitudes 2, 4, 6, and 8 km and only signs of E and S are considered. Thus, the output profile is classified into the 16 patterns with the binaries of the four layers.
The results show that the accuracy of prediction of the vertical profile pattern of E and S by the learned AI for the test data was about 20 ~ 25%. It was found that the AI parameterization is possible to roughly reproduce the geographical distribution of E and S in the one-month average as a climatological sense.