Use Observed Values In A Simulation​#

You can supply measured leaf area index (LAI) while calculating light interception. This is useful for testing the light model independently of a model that predicts LAI. Here LAI is m² leaf per m² ground, and incident PAR is an energy flux in W m⁻² ground.

One fixed observation​#

Supply a constant LAI in status, where the canopy stores its values, and use only the light model:

using PlantSimEngine, Dates, DataFrames
using PlantSimEngine.Examples

fixed_model = CompositeModel(
    Beer(0.6);
    status=(LAI=2.0,), id=:canopy, scale=:Canopy,
    environment=(Ri_PAR_f=100.0, duration=Day(1)),
)
fixed_simulation = run!(fixed_model; outputs=:all)
final_state(fixed_simulation).aPPFD

If you add a model that calculates LAI, it will replace this supplied value when it runs. To keep the observed LAI fixed, leave out that predictive model, or replace it with a model that supplies the observed value. In a composed scenario, the observation model should own the supplied output. Adding it alongside another writer of the same variable can create a writer conflict; Diagnostics.explain_writers(model) helps check ownership.

A sequence of observations​#

For observations that change over time, use a model that reads each observation and supplies it as LAI. The small ObservedLAI model below only copies the value; it performs no fitting, interpolation, or unit conversion. Include it from the downloadable source to run the example:

include("observed_lai.jl")

observations = DataFrame(
    duration=fill(Day(1), 3),
    measured_LAI=[1.0, 2.0, 3.0],
    Ri_PAR_f=fill(100.0, 3),
)
model = CompositeModel(
    ObservedLAI(), Beer(0.6);
    id=:canopy, scale=:Canopy, environment=observations,
)
simulation = run!(model; steps=nrow(observations), outputs=:all)
rows = collect_outputs(simulation; sink=DataFrame)
filter(row -> row.variable == :LAI || row.variable == :aPPFD, rows)

PlantSimEngine passes LAI from ObservedLAI to Beer automatically: one model supplies it and the other needs it on the same canopy. To predict LAI instead, replace ObservedLAI() with ToyLAIModel() and supply the thermal time that model requires. A change of model can change the inputs you must provide; compare alternatives before making a replacement.

The observation model​#

The complete source file is short:

struct ObservedLAI <: PlantSimEngine.Examples.AbstractLai_DynamicModel end

PlantSimEngine.inputs_(::ObservedLAI) = NamedTuple()
PlantSimEngine.outputs_(::ObservedLAI) = (LAI=0.0,)
PlantSimEngine.environment_inputs_(::ObservedLAI) = (measured_LAI=0.0,)

function PlantSimEngine.run!(::ObservedLAI, status, environment, constants, context)
    status.LAI = environment.measured_LAI
    return nothing
end

Match measurements to the simulation​#

Before using real measurements, check their source, units, whether they refer to leaf or ground area, their time stamps or measurement intervals, and any missing or rejected values. Associate measurements with stable object IDs; display names need not be unique. A plot-scale observation may also need an explicit aggregation or conversion before it can supply an organ-scale model.

This example has one observation per simulation day and one canopy. Its environment rows are already in simulation order; the observation model does not look up values by date or object ID. Prepare that alignment before using it. For several objects, arrange the observation source and bindings so each object receives its own measurement.

If measurements are less frequent than model execution, choose how to fill the gaps: for example, interpolate between measurements or keep the last measured value until the next one. State the interval over which each value applies and the policy for gaps or values outside the measured period. Document that choice before using the data. Filling a gap does not create a new independent observation for evaluation.

This approach supplies a measured input directly. It does not perform data assimilation, which would use observations to estimate or correct the model state and could account for measurement uncertainty. It also changes the question being tested: predicted light absorption under measured LAI checks the light component while treating canopy development as supplied data.

See Collect and plot results for retaining outputs and matching predictions with measurements, and Parameter fitting for calibration followed by independent evaluation. The observation-forcing prompt in Work with an AI coding agent can help you adapt this pattern to documented measurements and your intended objects.