Allow a constant, a computed constant, or a state variable (but not an algebraic variable) to be used as an initial value

[chrispbradley]

[And now in the right place!]

In the attached file (rename it .cellml), I attemt to initialise a state variable with a computed constant variable i.e., something along the lines of

var time: second {pub:in};
var state: second {pub: in};
var initial_value: dimensionless;
initial_value = 2.0{per_second}*state;
var state_variable: dimensionless {init: initial_value};
ode(state_variable,time) = 0.005{per_second} * state_variable;

This was previously possible in the old OpenCOR (and, as I understand it, you can also use state variables to initialise other state variables). However, in the new OpenCOR web app (version 0.20250714.0) I get the error

"Variable 'state_variable' in component 'test_component' is initialied using variable 'initial_value' which is not a constant."

It would be useful to allow state variables to by initialised by computed constant (and other state variables) variables as well as constant variables as in the old OpenCOR.

test_initial.txt

[agarny]

Thanks @chrispbradley. I agree that it should be possible to initialise a variable X with a variable Y as long as variable Y is either a constant, a computed constant, or a state variable (i.e. with an initial value of its own).

[nickerso]

is there a reason to rule out algebraic variables? assuming no circular dependencies between variables there is no actual restriction, right?

(just a question, not suggesting it needs to be included when addressing this issue!)

[agarny]

Algebraic variables are computed in computeRates() and/or in computeVariables() while the initialisation of a variable is done in either initialiseVariables() or computeComputedConstants(). Then, to compute a model, there are essentially two steps:

1. Call initialiseVariables(), computeComputedConstants(), computeRates(), and computeVariables() once to have everything initialised.
2. Call computeRates() and computeVariables() in a loop to "integrate" the model.

So, if we were to have a variable initialised using an algebraic variable, we would most likely need to initialise that variable in computeVariables(), but then that variable would always be initialised!

We could "fix" this by having another method (e.g., something like finaliseVariableInitialisations()) that would be called once at the end of the first step. This would however add another level of complexity for something that doesn't seem to have ever been requested in the past ~25 years?... (FWIW, the CellML API doesn't allow a variable to be initialised using an algebraic variable.)

In fact, to allow a variable to be initialised using a computed constant is already going to be a bit of a pain since computeComputedConstants() currently requires passing two arrays (constants and computedConstants) and we are now going to have to also pass states and algebraic. The former in case we want to initialise a state variable with a computed constantd while the latter in case we want to provide the initial guess of a non-linear algebraic variable (to be computed using KINSOL) using a computed constant. (I hadn't already thought of that latter case... man, a can of worms!)

So, bottom line: I would happily skip variable initialisation using an algebraic variable at this stage. :)

[agarny]

Talking of non-linear algebraic variables, we should allow them to be used to initialise another variable since they do have an initial value (which is used as the initial guess for KINSOL). Man... 🤦‍♂️

[agarny]

FTR, this has now been done (see PR #1325) and awaiting to be reviewed.