IN003-08
Harnessing HPC for Cloud Resolving NU-WRF Subseasonal Forecasts with AI Emulations
Monday, 7 December 2020: 07:21
Virtual
Milton Halem1, Jennifer Sleeman1, Zhifeng Yang2, Mian Chin3, Duncan Watson-Parris4 and Belay Demoz5, (1)University of Maryland Baltimore County, Computer Science, Baltimore, MD, United States, (2)University of Maryland Baltimore County, Physics Department, Baltimore, MD, United States, (3)NASA Goddard SFC, Greenbelt, MD, United States, (4)University of Oxford, Oxford, United Kingdom, (5)University of Maryland Baltimore County, Department of Physics & JCET, Baltimore, United States
Abstract:
In response to the Congressional Weather Research and Forecasting Innovation Act- Title II of 2017, NOAA engaged a multi-agency, multi-national North America Multi Model Ensemble (NMME) team to make subseasonal (32 days) seasonal global and regional forecasts on a weekly basis. The Congressional Act calls for improvements in the prediction of precipitation. This ensemble of global models currently employs spatial resolutions of ~ 25 km -100 km and several use microphysical and aerosol parameterizations that address the precipitation processes that occur on sub grid scales. sub-grid microphysics schemes parameterize cloud condensation, sublimation, evaporation and sedimentation of liquid and ice based on nuclei condensates derived from the aerosol module parameterizations. The microphysics parameterizations have differing numbers of moisture variables and represent water, ice, and mixed-phase processes that result from the interaction of ice and water particles. The NCAR WRF model developer team recommends using these mixed-phase processes for grid sizes less than 10 km, particularly in convective or icing situations but for coarser grids they claim it may not be worth the added expense of these schemes.
We will present the results of a cloud resolving subseasonal regional forecast over the CONUS employing the NU-WRF model system at a spatial resolution of 4km and 64 vertical levels with two prescribed lateral boundary conditions (LBC), one from MERRA and the other from a GEOS-5 forecast. The NU-WRF system includes the GOCART aerosol, a two moment bulk microphysics parameterization and excludes chemistry, photochemistry and land processes. Since the microphysics and GOCART parameterization consume more than 40% of the compute resources, we explored replacing the aerosol and microphysics parameterizations with machine learning emulators. We employ distributed implementations of Amoeba Net and DENSE Net for off-line training of the aerosol and microphysics parameterizations and will present the results.
This study is being planned for execution on the NASA NCCS computer system over a 4 month period at 2 hours /sim/day. Finally, we will assess the potential for a coupled global 10 km model for LBC and 1 km spatial NU-WRF model to harness a future Exascale system.
