Oct WRF 4D-Var System Xiang-Yu Huang, Xin Zhang Qingnong Xiao, Zaizhong Ma, John Michalakes, Tom henderson and Wei Huang MMM Division National Center for Atmospheric Research Boulder, Colorado, USA
Oct Contents Current status of WRF 4D-Var Scientific Performance of WRF 4D-Var Software Engineering Performance of WRF 4D-Var On-going Works
Oct (old forecast) (new) (initial condition for NWP) xx 4D Variational Data Assimilation
Oct J' vn = v n + v i + U T S V-W T M k T S W-V T H k T R -1 [H k S W-V M k S V-W U -1 v n + H k (M k (x n-1 )) – y k ] n-1 i=1 K k=1 Black – WRF-3DVar [B, R, U=B 1/2, v n =U -1 (x n -x n-1 )] Green – modification required Blue– existing (for 4DVar)--WRF Red – new development (Huang, et.al. 2006: Preliminary results of WRF 4D-Var. WRF users’ workshop, Boulder, Colorado.) Current Status of WRF 4D-Var WRF AD WRF TL ARW WRF
Oct Basic system: 3 exes, disk I/O, parallel, full dyn, simple phys, JcDF
Oct Single Observation Experiments
Oct
8
9
10
Oct Multi-incremental 4D-Var (Adopted from The Operational Data Assimilation System, Lars Isaksen, ECMWF)
Oct Haitang case Outerloop 15km, 271x217x17, time step: 90s Innerloops 45km, 91x73x17, time step: 240s 135km, 31x25x17, time step :600s
Oct Convert WRF-Var on A-grid to WRF on C-grid (By Rizvi) VVVV UXUXUXUXUVVVVUXUXUXUX UVVVV UXUXUXUX U VVVV UXUXUXUX UVVVVVVVV UXUXUXUXUVVVVUXUXUXUX UVVVV UXUXUXUX U VVVV UXUXUXUX UVVVV
Oct What is needed for C-Grid? Observation operators, its tangent linear and adjoint Computation of innovations Computation of gradients PSI-CHI to UV solver and its adjoint Tackle Halo region for MPI run
Oct Analysis difference (A2C - trunk) Only GPSPW Obs. Date: Date: CONUS 200 KM T8 15 KM
Oct Scientific Performance of WRF 4D-Var Typhoon Haitang experiments: 5 experiments, every 6 h, 00Z 16 July - 00 Z 18 July, Typhoon Haitang hit Taiwan 00Z 18 July FGS – forecast from the background [The background fields are 6-h WRF forecasts from NCEP GFS analysis.] 2.AVN – forecast from the NCEP GFS analysis 3.3DVAR – forecast from WRF 3D-Var 4.FGAT – first guess at appropriate time ( A option of WRF-3DVAR) 5.4DVAR – forecast from WRF 4D-Var Domain size: 91x73x17 Resolution: 45 km Time Window: 6 Hours, Observations: GTS conventional observations, bogus data from CWB
Oct Typhoon Haitang Verification 48-h forecast typhoon tracks from FGS, AVN, 3DVAR, FGAT, 4DVAR, together with the observed best track. Forecasts are a ll started from 0000 UTC 16 July 2005.
Oct KMA Heavy Rain Case Period: 12 UTC 4 May - 00 UTC 9 May, 2006 Grid : (60,54,31) Resolution : 30km Domain size: the same as the operational 10km do-main. Assimilation window: 6 hours Warm started cycling run
Oct Precipitation Verification
Oct T8 domain case study Period : Size: 123x111x27 Resolution: 45km Experiment design: 3/6 hours GFS forecast as FG 1. 3DVAR 2. 4DVAR cycling run, 24h forecast after analysis Observations used: soundsonde_sfc synopgeoamv aireppilot ships qscat gpspw gpsref
Oct Track and Intensity Forecast 3D-Var 4D-Var
Oct The radar data assimilation experiment using WRF 4D-Var Yong-Run Guo and Jenny Sun TRUTH Initial condition from TRUTH (13-h forecast initialized at Z from AWIPS 3-h analysis) run cutted by ndown, boundary condition from NCEP GFS data. GFS Both initial condition and boundary condition from NCEP GFS data. 3DVAR DVAR analysis at Z used as the initial condition, and boundary condition from NCEP GFS. Only Radar radial velocity at Z assimilated (total # of data points = 65,195). 4DVAR DVAR analysis at Z used as initial condition, and boundary condition from NCEP GFS. The radar radial velocity at 4 times: , 05, 10, and 15, are assimilated (total # of data points = 262,445).
Oct TRUTHGFS4DVAR3DVAR Hourly precipitation ending at 03-h forecast
Oct TRUTH GFS3DVAR4DVAR Hourly precipitation ending at 06-h forecast
Oct Real data experiments
Oct For general cases, the performance of WRF 4D-Var is comparable with WRF 3D-Var. For some fast developing, fine scale cases such as squall line, tropical cyclone, heavy rainfall case, WRF 4D-Var does a much better job than 3D- Var.
Oct Software Engineering Performance of WRF 4D-Var Ability to assimilation all kinds of observation as 3D-Var (Radiance and Radar). Both serial and parallel runs are supported. Tested Platforms: IBM with XLF, Linux with PGI,INTEL & G95, Mac G5 with G95 & XLF. Multi-incremental 4D-Var. Flexible assimilation time window (for example, 15 minutes ~ 6 hours)
Oct Preliminary Disk IO Optimization-Case Study Trade memory for Disk IO in 4D-Var Domain size : 151x118x31 Resolution :4km Time-step : 25s Time window :15m # of iterations : 60 Obs. : OSSE radar wind # of obs. : In this case, 82% disk IO was replaced by memory IO
Oct On-going Works Remove Disk IO which is used as communication among WRF 4D-Var components, ESMF is a candidate.(~50% wall-clock time reduction, improve parallel scalability) Cleanup solve_em_ad (~90% cost), trade re-computation with memory (another ~50% wall-clock time reduction).