MSC_repo

Table of Contents

1. Project introduction
2. Used grids
3. Input data description
   • Files list
   • Variables
4. Machine learning
   • Tree regression

Project introduction

Topic: Use machine learning to revisit the data and questions of Griewank, Schemann, Neggers 2018
ICON high resolution (156m) is beneficial to the representation of processes that are subgrid-scale for regular (horizontal resolution of 2.8 km for COSMO) numerical weather forecast and climate models.
PDF cloud scheme parameterizes thermodynamic variables which ICON-LES resolves explicitly. Thus we can compare parametrization with "true" data - 3D high resolution ICON-LES data. The idea of Griewank, Schemann, Neggers 2018 article was to tweak different parameters and degrees of freedom of PDF cloud scheme, also change the grid cell the PDF applied to. These changes produce different results and then compared to "true" distributions of td variables from ICON-LES high resolved data. Also performances of skewed and symmetrical PDFs were compared.

Used grids

ICON - LES:

• *horizontal resolution:156-m triangle edge grid cell
• vertical resolution:dz changing from 20m in the near-surface area to the 1000 m at the top.
• n vertical layers:150, n=150 - surface, n=1 - top of the atmosphere.
• cell shape: triangle
  *resolution refers to the square root of the mean cell area in the icosahedral grid of ICON
  

CLOUD FRACTION PARAMETRIZATION (Griewank, Schemann, Neggers 2018):

• horizontal resolutions:
  • 110 × 110 km,20 subdomains inside cut out area
  • 55 × 55 km, 80 subdomains inside cut out area
  • 27.5 x 27.5 km, 320 subdomains inside cut out area
  • 13.8 x 13.8 km, 1280 subdomains inside cut out area
  • 7 x 7 km, 5120 subdomains inside cut out area
• vertical resolution:?
• n vertical layers:150, n=150 - surface, n=1 - top of the atmosphere
• cell shape: square longitude latitude subdomains

NETCDF data description

Files list

"douze" stands for 12; douze files contain all the values from all 12 snapshots indepenetly
Grid cells called SUBDOMAINS below

ICON-LEM has 156-m triangle edge grid cells
NT=12,12 3-D snapshots simulated by the ICON model in LES mode at 12 point in time:

• 24 April 2013 12:35, 18:35;
  
• 25 April 2013 12:07, 14:07;
  
• 26 April 2013 18:22,
  
• 11 May 2013 14:06;
  
• 29 July 2014 06:00, 14:00;
  
• 14 August 2014 12:03 18:03
  
• 3 June 2016 6:00, 14:00.
  
  

NETCDF Variables

INPUTS FROM LES
douze files contain 63 variables, among them such variables as:

• qvlm : Mean liquid water + vapor mixing ratio (liquid and gas water content) [kg/kg]
• qvl_qs : the saturation deficit when qvlm is smaller than qsm ??
• qvlm - qsm is the saturation deficit when qvlm is smaller than qsm
• qcm : Mean cloud water mixing ratio [kg/kg]
• qsm : Saturation mixing ratio [kg/kg] - ratio of the mass of water vapor to the mass of dry air in a parcel of air at saturation.
• qvm : Mean water vapour mixing ratio[kg/kg] - ratio of the mass of water vapour to the mass of dry air
• qlm : Liquid water mixing ratio [kg/kg] - the ratio of the mass of liquid water to the mass of dry air in a unit volume of air. Units are mass of liquid water per mass of dry air
• skew : LES water vapor skewness
• skew_l : LES liquid water+water vapor skewness
• var_l : variance of Liquid water + vapor
• var_t : temperature variance
• pm : Mean pressure [Pa]
• tm : Mean temperature [K]
• zm : Mean height [m] Values from 0 to 21000
  Naming system: q means water content, v is vapor, l is liquid, m is mean. so qvlm is the liquid and gas water content.
  *mean above refers to spatial averaging of ICON-LES gridcells within particular "Cloud fraction parametrization" subdomain?
  

Below are plots of some variable from ncr_pdf_douze_1deg.nc.
There are 20 horizontal subdomains of size 110 × 110 km for each of 12 snapshots. Thus 20*12=240 subdomain columns - Y AXIS. Each of these subdomain columns has nz=150 values of each variable, such as pm. These values were calculated over all the LES gridcells(156m size) in the subdomain.

CLOUD FRACTIONS
cl_l : LES liquid* cloud fraction, determined from the number of saturated cells(156m) divided by the total number of cells per slice(subdomain)
cl_msl : cloud fraction of parametrization V "mean saturation" closure
cl_zsl : cloud fraction of parametrization VI "zero buoyancy" closure
cl_rel : cloud fraction of parametrization II - "relative humidity" closure
cl_zff : cloud fraction of parametrization VIII - no closure
*only liquid clouds are considered in liquid c.fraction as opposed to ice cloud fraction. Ice clouds are not taken into account for simplicity.

Machine learning

Tree regression

input variables:

From 63 variables of douze files few variables below were selected as an ML alrgorithm input.

• 'qsm', 'qtm', 'qlm', 'skew_l', 'var_l', 'var_t', 'tm', 'pm' - ICON-LES values averaged within the corresponding "cloud parametrization" subdomain.

These varaibles were packed in the input vector of shape: ([N_snapshots * N_subdomains * N_vertical_levels] X N input variables)
Below the histograms of input variables for ncr_pdf_douze_1deg.nc and ncr_pdf_douze_05deg.nc are shown:

Spatial averaging influences distributions of IO variables as one can see below:
Axis are be set to a log scale to better see the difference between the various scales.

goal variables:

• cl_l : liquid cloud fraction of LES, determined from the number of saturated cells(156m) divided by the total number of cells per slice(subdomain)
  
• vector shape: goal vairable is a vector of shape ([N_snapshots * N_subdomains * N_vertical_levels] X 1)

Input are variables split into:

• test(evaluation) fraction - fraction of input/goal variables for evaluation of regression method
• training fraction= 1- test fraction (the rest of the data)