New API

[JordiBolibar]

We should improve the interface of how we launch simulations. In order to have everything in a more tidy way. We should create data structures that gather multiple parameters driving simulations.

Here's a draft of the prototype I have in mind:

Model parameters

• PhysicalParameters struct to hold all parameters related to model physics.

• Hyperparameters struct holding all the hyperparameters from the neural network.

• SolverParameters struct holding all parameters to configure the solver.

• UDEparameters struct holding all parameters for the training of the UDE. It's a child of SolverParameters, since it also includes a solver.

• OGGMparameters struct wrapping OGGM's parameters

• SimulationParameters struct holding the parameters defining a simulation

• Parameters: struct holding all the parameter types.

Models

• Machine struct holding an ML model (e.g. a neural network) + a Hyperparameters struct. Inspired by MLJ.jl. When initialized inside a Model, it automatically fetches the Hyperparameters to generate the NN.

• MBmodel struct defining a mass balance model. It can have children like TImodel.

• IceflowModel struct defining an ice flow model. It can have children, like SIAmodel

• Model struct, contains all information about the PhysicalParameters, SolverParameters, UDEparameters, OGGMparameters, Machine, MBmodel, IceflowModel

Glaciers, intial conditions and climate data

• Climate struct containing climate data series for a Glacier.

• Glaciers struct containing OGGM's gdirs and additional information on the Julia side, for example its Climate.

Simulations

• Simulation (inmutable?) struct holding all the data necessary for a simulation: Model, Glaciers, its type (functional inversion, inversion or forward).

To be run such as:

# Produce everything needed for a simulation
# First we create the model parameters
parameters = Parameters(; physics = PhysicsParameters(),
                                      hyper = Hyperparameters(),
                                      solver = SolverParameters(),
                                      UDE = UDEparameters(),
                                      OGGM = OGGMparameters(workspace_path),
                                      simulation = SimulationParameters())

# Then we generate the models to be run
model = Model(IceflowModel(), MBmodel(), Machine(), params)
glaciers = Glaciers(rgi_ids)
add_climate!(Glaciers, time_period)

# We create a simulation based on everything we generated before
simulation = Simulation(model, glaciers, parameters)

# We finally run the simulation
run!(simulation)

Beyond this, we should also accommodate an API similar to OGGM's, where everything is based on entity_tasks around glacier directories. Since Julia is not OO, entity tasks would just be functions with a common interface, run on Glaciers. Something like:

entity_tasks = [ODINN.tasks.DoSomethingToGlaciers(), ODINN.tasks.DoSomethingElseToGlaciers()]
apply_entity_tasks!(glaciers, entity_tasks)