Parameter fitting

The fit method

Models are often calibrated using data, but the calibration process is not always the same depending on the model, and the data available to the user.

PlantSimEngine defines a generic fit function that allows modelers provide a fitting algorithm for their model, and for users to use this method to calibrate the model using data.

The function does nothing in this package, it is only defined to provide a common interface for all the models. It is up to the modeler to implement the method for their model.

The method is implemented as a function with the following design pattern: the call to the function should take the model type as the first argument (T::Type{<:AbstractModel}), the data as the second argument (as a Table.jl compatible type, such as DataFrame), and any more information as keyword arguments, e.g. constants or parameters initializations with default values when necessary.

Example with Beer

The example script (see src/examples/Beer.jl) that implements the Beer model provides an example of how to implement the fit method for a model:

function PlantSimEngine.fit(::Type{Beer}, df; J_to_umol=PlantMeteo.Constants().J_to_umol)
    k = Statistics.mean(log.(df.Ri_PAR_f ./ (df.PPFD ./ J_to_umol)) ./ df.LAI)
    return (k=k,)
end

The function takes a Beer type as the first argument, the data as a Tables.jl compatible type, such as a DataFrame as the second argument, and the J_to_umol constant as a keyword argument, which is used to convert between μ mol m⁻² s⁻¹ and J m⁻² s⁻¹.

df should contain the columns PPFD (μ mol m⁻² s⁻¹), LAI (m² m⁻²) and Ri_PAR_f (W m⁻²). The function then computes k based on these values, and returns it as a NamedTuple of the form (parameter_name=parameter_value,).

Here's an example of how to use the fit method:

Importing the script first:

using PlantSimEngine, PlantMeteo, DataFrames, Statistics
# Import the examples defined in the `Examples` sub-module:
using PlantSimEngine.Examples

Defining the meteo data:

meteo = Atmosphere(T=20.0, Wind=1.0, P=101.3, Rh=0.65, Ri_PAR_f=300.0)

Atmosphere(date = Dates.DateTime("2024-10-28T16:50:01.477"), duration = Dates.Second(1), T = 20.0, Wind = 1.0, P = 101.3, Rh = 0.65, Precipitations = 0.0, Cₐ = 400.0, e = 1.5255470730405223, eₛ = 2.3469954969854188, VPD = 0.8214484239448965, ρ = 1.2037851579511918, λ = 2.4537e6, γ = 0.06723680111943287, ε = 0.5848056484857892, Δ = 0.14573378083416522, clearness = Inf, Ri_SW_f = Inf, Ri_PAR_f = 300.0, Ri_NIR_f = Inf, Ri_TIR_f = Inf, Ri_custom_f = Inf)

Computing the PPFD values from the Ri_PAR_f values using the Beer model (with k=0.6):

m = ModelList(Beer(0.6), status=(LAI=2.0,))
run!(m, meteo)

Now we can define the "data" to fit the model using the simulated PPFD values:

df = DataFrame(aPPFD=m[:aPPFD][1], LAI=m.status.LAI[1], Ri_PAR_f=meteo.Ri_PAR_f[1])

And finally we can fit the model using the fit method:

fit(Beer, df)

(k = 0.6,)