Implement ClimaLand global calibration

[AlexisRenchon]

Tasks

Beta Give feedback

1. 1 column, perfect model (done, link PR) Add a calibration tutorial for single site perfect model  #621
2. 1 column, observation, make sure prior sample is a large spread around target! (to do so, 1, identify correct parameters, and 2, their correct distribution - so that wide spread around target) Fix calibration perfect model, add observations tutorial, fix ClimaLandSimulations #816
3. Global simulation (100s columns) with global bucket, LHF and SHF (ERA5), monthly, 1 year (multiple obs., multiple locations, write read climadiagnostics files, uses EKP Observations) Calibrate global bucket LHF and SHF #835
4. Global simulation with land model

[AlexisRenchon]

• Use ETKI, see https://clima.github.io/EnsembleKalmanProcesses.jl/dev/ensemble_kalman_inversion/

• Multicolumn using EKP Observation objects, see https://clima.github.io/EnsembleKalmanProcesses.jl/dev/observations/

Observations:

• first, one per variable (same location), e.g., LHF, GPP with combine_observation
• Then, combine them for multiple locations with Observationseries

[AlexisRenchon]

Data (LHF and SHF that we calibrate on) needs to be in the same format as model output

[AlexisRenchon]

Note: we should change single site observation calibration a bit, it takes 2 hours right now (300 simulations run), we could save the figure but run it with less simulations to ensure it passes CI

[AlexisRenchon]

Land model:
What loss function?
LHF, SHF, emissivity, SW, LW, (SIF, NEE, C fluxes)

[AlexisRenchon]

• Format: EKP vs. ClimaCalibrate
• Models: Bucket vs. LandModel
• Data: Observations vs. Perfect Model
• Scale: Single site vs. Multi-columns vs. Global

Possible KR:

• Create pipeline with EKP or ClimaCalibrate
• Demonstrate parameters recovery with Perfect Model
• Demonstrate X with observations

What we mean by Loss:

• Different averaging period or other reduction of variables
• Different configuration, e.g., initialisation

Parameters:

• Correlations between parameters
• Parameters selection and distributions

Next to do:

• Multiple variables stacked (ETKI, see comment above)
• Multiple columns (e.g., 100s, see above)
• Using netcdf diagnostics read and write

What we can qualify / quantify (and remember, stochastic method -> change all the time):

• PM: collapse of ensemble toward original parameters
• prior spread vs. final iteration spread
• correlation between params

random notes:

• Julia does not preserve seed between version!

[AlexisRenchon]

(to do: add 2*sqrt(noise) as a band around target)

[AlexisRenchon]

Options: limits the amount of data by averaging over larger steps

[AlexisRenchon]

Think of informative features of the data that you want to capture such as seasonal amplitude, seasonal shifts, diurnal amplitude, etc. that are controlled by parameters...

Then you can make THAT your target as a function of params, and calibrate

[AlexisRenchon]

observational noise is standard deviation, i.e., covariance

[AlexisRenchon]

next steps:

• ClimaCalibrate for Bucket
• g(params) for full land model, global calibration

[AlexisRenchon]

when doing random locations on land, do we sample good representation of all landscapes

[AlexisRenchon]

only write LHF and SHF diagnostics