ICON The new global nonhydrostatic model of DWD
ICON The new global nonhydrostatic model of DWD and MPI-M Daniel Reinert 1, Günther Zängl 1, and the ICON-team 1, 2 1 Deutscher Wetterdienst / 2 Max-Planck-Institute for Meteorology 13 th EMS Annual Meeting 09 – 13 September 2013, Reading, United Kingdom
ICON – ICOsahedral Nonhydrostatic Model Ø Joint development project of DWD and Max-Planck. Institute for Meteorology for building a nextgeneration global NWP and climate modelling system MPI DWD Ø Atmosphere and ocean model Outline I. Project goals II. Horizontal grid structure and accompanying problems III. ICON NWP physics suite IV. Selected results V. Roadmap and Summary Daniel Reinert – 12. 09. 2013 2
I. Primary development goals Ø Improved conservation properties (at least mass) and consistent tracer transport (tracer air-mass consistency) Ø Applicability on a wide range of scales from 100 km to 1 km Ø Scalability and efficiency on massively parallel computer architectures with O(104 +) cores Ø Local refinement/nesting capability Ø At DWD: • Replace current global model GME • Replace regional model COSMO-EU by a highresolution window over Europe. Ø At MPI-M: • Use ICON as dynamical core of an Earth System Model (MPI-ESM 2) Daniel Reinert – 12. 09. 2013 Horizontal grid with nest over Europe 3
II. ICON’s unstructured grid Primal cells: triangles Ø uses icosahedron for macro triangulation Ø C-type staggering: § velocity at edge midpoints § mass at cell circumcenter Ø local subdomains (“nests”) Triangular C-Grid local domain(s) global domain Daniel Reinert – 12. 09. 2013 4
Equations (dry adiabatic) and solver Ø Fully compressible nonhydrostatic vector invariant form, shallow atm. Solver: Ø Finite volume/finite difference discretization (mostly 2 nd order) Ø Two-time level predictor-corrector time integration Ø Vertically implicit (vertical sound-wave propagation) Ø Fully explicit time integration in the horizontal (at sound wave time step; not split explicit!) Ø Mass conserving Daniel Reinert – 12. 09. 2013 5
Checkerboard noise on triangular C-Grid è Main problem with triangular C-grid: suffers from spurious computational mode (e. g. Danilov (2010)), triggered by the discretized divergence operator (Wan (2013)) è Divergence operator: applies the Gauss theorem è Truncation error (Wan (2013)): Ø Only 1 st order accurate on triangular C-grid Ø Error changes sign from upward- to downward pointing triangle checkerboard Example for synthetic velocity field (Wan, 2013) Daniel Reinert – 12. 09. 2013 6
Controlling the checkerboard noise è Goal: Eliminate 1 st order error è Basic idea: Divergence averaging I: Compute standard 1 st II: Compute divergence estimate based on immediate neighbors (2 nd order bilinear interpolation) order div III Averaging: 2 nd order accurate for isosceles triangles Daniel Reinert – 12. 09. 2013 7
Example: Baroclinic wave Jablonowski-Williamson (2006) baroclinic wave test case PS T Daniel Reinert – 12. 09. 2013 8
Example: Baroclinic wave Jablonowski-Williamson (2006) baroclinic wave test case T PS Standard divergence operator Divergence averaging div “checkerboard” noise Daniel Reinert – 12. 09. 2013 9
III. ICON NWP-physics Process Author Scheme Origin Radiation Mlawer et al. (1997) Barker et al. (2002) RRTM ECHAM 6 Non-orographic gravity wave drag Scinocca (2003) Orr, Bechthold et al. (2010) wave dissipation at critical level IFS Cloud cover Köhler et al. (new development) diagnostic (later prognostic) PDF ICON Microphysics Doms and Schättler (2004) Seifert (2010) prognostic: water vapour, cloud water, cloud ice, rain, snow COSMO Saturation adjustment Blahak (2010) isochoric adjustment COSMO Convection Tiedtke (1989) Bechthold et al. (2008) mass-flux shallow and deep IFS Sub-grid scale orographic drag Lott and Miller (1997) blocking, GWD IFS Turbulent transfer / diffusion Raschendorfer (2001) prognostic TKE COSMO Soil/surface Heise and Schrodin (2002) Mironov and Ritter (2004) Mironov (2008) TERRA (tiled + multi-layer snow) SEAICE FLAKE(fresh water lake scheme) GME/COSMO Daniel Reinert – 12. 09. 2013 10
Reduced grid for radiation Ø Hierarchical structure of the triangular mesh is very attractive for calculating physical processes (e. g. radiative transfer) with different spatial resolution compared to dynamics. Radiation step every 30 min Radiative transfer computations Empirical corrections upscaling downscaling Daniel Reinert – 12. 09. 2013 11
Proof of concept Ø net surface shortwave flux (reduced – full grid) Ø average over 30 x 48 h forecast runs in June 2012 Avg: 1. 57 Reduced radiation grid currently generates positive bias in Daniel Reinert – 12. 09. 2013 12
Flat-MPI performance time (s) Recall goal: scalability up to O(104+) cores MPI tasks 1024 4096 Test setup: ICON RAPS 2. 0, IBM Power 7 20/10/5 km, 8 h forecast, reduced radiation grid (S. Körner, DWD, 03/2013) Daniel Reinert – 12. 09. 2013 13
IV. Selected results of NWP test suite è Real-case 7 -day forecasts with interpolated IFS analysis data è WMO standard verification against IFS analysis on 1. 5° lat/lon grid. è Comparison against GME reference experiment with interpolated IFS analysis data. ICON 40 L 90 GME 40 L 60 hor. resolution 40 km vertical levels 90 60 top height 75 km 36 km analysis data IFS Basic requirement for operational use of ICON must outperform GME in terms of forecast quality/scores Daniel Reinert – 12. 09. 2013 14
Verification: Surface Pressure, January 2012 Region: Northern hemisphere (NH) ICON GME against IFS SH: 21% Verification: G. Zängl, U. Damrath, 08/2013 (DWD) Daniel Reinert – 12. 09. 2013 15
Verification: Geopot 500 h. Pa, January 2012 Region: Northern hemisphere (NH) ICON GME against IFS SH: 9. 4% Verification: G. Zängl, U. Damrath, 08/2013 (DWD) Daniel Reinert – 12. 09. 2013 16
Verification: Rh 700 h. Pa, January 2012 Region: Tropics (Tr) ICON GME against IFS ICON shows strong positive moisture bias in the tropics Verification: G. Zängl, U. Damrath, 08/2013 (DWD) Daniel Reinert – 12. 09. 2013 17
V. Roadmap towards operational application Daniel Reinert – 12. 09. 2013 18
Summary ICON is entering the home stretch for becoming operational Ø Verification results are mostly exceeding those of GME, but there are still some weaknesses/biases e. g. moisture field Ø Technical parts scale on massively parallel systems (I/O still needs performance improvements) Ø Optimization of forecast quality still ongoing Ø Tests with own 3 D-Var data assimilation have started recently. Daniel Reinert – 12. 09. 2013 19
Thank you for your attention !!
- Slides: 20