Surface fluxes of heat, moisture and momentum: Land and ocean seasonal climatology

[milankl]

This implements surface fluxes of heat, moisture and momentum following Fortran SPEEDY but with shortcuts (no skin temperature calculation for now). Land and ocean temperatures are seasonal climatologies updated every 3 days.

[milankl]

Surface humidity 1 day after starting from a dry atmosphere, fluxes currently not based on skin temperatures, vegetation ignored (availability = 0.2 everywhere)

10 days in

30 days in

still not high enough to precipitate.

@p-hss

[milankl]

Surface specific humidity increase over 4 months (time vs lat @ 180E)

And during the 8 months after that

reaching still higher humidity levels, but again not enough to precipitate by itself

[milankl]

land_surface_temperature.nc from previous land.nc created via (to include a mask and compression)

using NCDatasets, Dates, BitInformation
ds = NCDataset("land.nc")
ds2 = NCDataset("land_surface_temperature.nc","c")
nx, ny, nt = ds.dim["lon"], ds.dim["lat"], ds.dim["time"]

defDim(ds2,"lon",nx)
defDim(ds2,"lat",ny)
defDim(ds2,"time",nt)
t0 = Dates.DateTime(1981,1,1)

lon = defVar(ds2,"lon",Float32,("lon",),attrib=Dict("units"=>"degrees_east","long_name"=>"longitude"))
lat = defVar(ds2,"lat",Float32,("lat",),attrib=Dict("units"=>"degrees_north","long_name"=>"latitude"))
time = defVar(ds2,"time",Int64,("time",),attrib=Dict("units"=>"hours since 1981-1-1 00:00:00",
                                                    "standard_name"=>"time",
                                                    "calendar"=>"gregorian"))

lon[:] = ds["lon"][:]
lat[:] = ds["lat"][:]

for t in 1:nt
    time[t] = Dates.Hour.(ds["time"][t] - t0).value
end

lst = defVar(ds2,"lst",Float32,("lon","lat","time"),
attrib=Dict("long_name"=>"land surface temperature",
            "units"=>"K","_FillValue"=>ds["stl"][1,1,1]),
deflatelevel=9,shuffle=true)

keepbits = 11
lst[:] = round(ds["stl"][:],keepbits)

close(ds2)

[milankl]

Last commit brings the model initialisation time down to 0.1-0.2s (was 2-4s).

Before: (no ocean or land initialization, which is an interpolation for every month)

julia> @time simulation = initialize!(model)
  1.691842 seconds (437.42 k allocations: 108.633 MiB, 1.29% gc time)
SpeedyWeather.Simulation{PrimitiveWetModel{Float32, SpeedyWeather.CPUDevice}}
├ PrimitiveWetModel{Float32, SpeedyWeather.CPUDevice}
├ PrognosticVariables{Float32, OctahedralGaussianGrid{Float32}, PrimitiveWetModel{Float32, SpeedyWeather.CPUDevice}}
└ DiagnosticVariables{Float32, OctahedralGaussianGrid{Float32}, PrimitiveWet}

After: (with ocean, land and orography)

julia> @time simulation = initialize!(model)
  0.137597 seconds (641.46 k allocations: 61.707 MiB)
Main.SpeedyWeather.Simulation{PrimitiveWetModel{Float32, Main.SpeedyWeather.CPUDevice}}
├ PrimitiveWetModel{Float32, Main.SpeedyWeather.CPUDevice}
├ PrognosticVariables{Float32, OctahedralGaussianGrid{Float32}, PrimitiveWet}
└ DiagnosticVariables{Float32, OctahedralGaussianGrid{Float32}, PrimitiveWet}

Orography had an expensive large spectal transform (interpolation now), ocean and land read individual values from netcdf (slow, now in batch).