README

README concerning the extract of the coarse-grained NARVAL and QUBICC data available via DKRZ resources 

Date: 9/1/2022

T I T L E O F T H E D A T A S E T
- NARVAL R2B10 coarse-grained to R2B4
- NARVAL R2B10 coarse-grained to R2B5
- QUBICC R2B9 coarse-grained to R2B5

P A T H T O T H E D A T A S E T
- /home/b/b309170/workspace_icon-ml/iconml_clc/extract_from_the_data

O W N E R / P R O D U C E R O F T H E D A T A S E T 
- NARVAL R2B10: Daniel Klocke and Deutscher Wetterdienst, Offenbach and Max-Planck-Institut für Meteorologie, Hamburg
Coarse-grained to R2B4 and R2B5: Arthur Grundner and Deutsches Zentrum fuer Luft- und Raumfahrt e. V., Institut für Physik der Atmosphäre, Oberpfaffenhofen
- QUBICC R2B9: Marco Giorgetta and Max-Planck-Institut für Meteorologie, Hamburg 
Coarse-grained to R2B5: Arthur Grundner and Deutsches Zentrum fuer Luft- und Raumfahrt e. V., Institut für Physik der Atmosphäre, Oberpfaffenhofen

D A T A U S A G E L I C E N S E 
- See LICENSE File in this directory

C O N T E N T O F T H E D A T A S E T
- To an R2B4 and an R2B5 grid coarse-grained data from the NARVAL (Klocke et al., 2017: 'Rediscovery of the doldrums in storm-resolving simulations over the tropical Atlantic') 
  and the QUBICC hindcast simulations hc2, hc3, hc4 (Giorgetta et al., 2022: 'The ICON-A model for direct QBO simulations on GPUs')
- The dataset contains one timestep per day of cloud volume fraction data and one timestep with all variables
- fr_lake: Fraction of lakes (2D), fr_land: Fraction of land (2D), pres/pfull: Pressure [Pa], qc/clw: Specific cloud water content [kg/kg], 
  qi/cli: Specific cloud ice content [kg/kg], qv/hus: Specific humidity [kg/kg], rho: Air density [kg/m^3], temp/ta: Air temperature [K], u/ua: Zonal wind [m/s], 
  v/va: Meridional wind [m/s], zg: Geometric height at full levels [m], zghalf: Geometric height at half levels [m], clc/cl: Cloud volume fraction [%],
  cl_area: Cloud area fraction [%]

D A T A U S A G E S C E N A R I O S
- To study the link between the thermodynamic environment and cloud cover using the output of storm-resolving model simulations on a climate model resolution
- To train ML-based parameterizations

M E T H O D S U S E D F O R D A T A C R E A T I O N
- See 'Deep Learning Based Cloud Cover Parameterization for ICON' (Grundner et al., 2022)

I S S U E S
- There is a temporal mismatch between some model output variables from one common time step. This is caused by the sequential processing of some parameterization schemes 
  in the ICON model. For instance, the cloud cover scheme diagnoses cloud cover before the microphysics scheme alters the cloud condensate mixing ratio, which has led 
  to roughly 7% of the cloudy grid cells in our data to be condensate-free. However, this mismatch should not exceed the fast physics time step in the model, which 
  was set to 40 seconds in the QUBICC and to 24 seconds in the NARVAL simulations.
- The mixing length in the vertical diffusion scheme was mistakenly set to 1000m instead of 150m, causing unrealistically strong vertical diffusion in some situations.