Unnecessary mixed-integer equations could be solved by better BLT

[Francesco Casella]

Consider the following test cases (the second requires to install the ExternalMedia library from ​https://github.com/modelica/ExternalMedia, I am sorry but I could not replicate the problem in a simpler way)

package TestMixedInteger
  package TestMedium
    extends Modelica.Media.Interfaces.PartialTwoPhaseMedium;

    redeclare record extends ThermodynamicState
      Temperature T;
      AbsolutePressure p;
      SpecificEnthalpy h;
      Density d;
    end ThermodynamicState;

    redeclare function setState_ph
      input AbsolutePressure p;
      input SpecificEnthalpy h;
      input FixedPhase phase = 0;
      output ThermodynamicState state;
    algorithm
      state.p := p;
      state.h := h;
      state.T := (h-1e5)/4200 + 20;
      state.d := 995 - (state.T-293)*0.1;
      state.phase :=0;
    end setState_ph;

    redeclare function extends density
    algorithm
      d := state.d;
    end density;

  end TestMedium;

  model Test1
    replaceable package Medium = TestMedium
      constrainedby Modelica.Media.Interfaces.PartialTwoPhaseMedium;
    Medium.AbsolutePressure p1(start = 6e5);
    Medium.Density rho;
    Medium.ThermodynamicState state;
    Modelica.SIunits.MassFlowRate m_flow;
  equation
    state = Medium.setState_ph(p1,1e5);
    rho = Medium.density(state);
    m_flow = 1e-4*sqrt(rho*(p1-5e5));
    m_flow = 1;
  end Test1;

  model Test2
    extends Test1(redeclare package Medium =
          ExternalMedia.Examples.WaterCoolProp);
  end Test2;

  model Test3
    extends Test1(redeclare package Medium =  
          Modelica.Media.Water.StandardWater);
  end Test3;
end TestMixedInteger;

Code generated by OMC r22383 for Test1 and Test3 do not include any mixed-integer equations. Test2 instead ends up in a mixed-integer system of equations, involving the state.phase Integer.

This is completely unnecessary, as Test2 could be solved this way:

m_flow = 1;
// solved as m_flow :=1

{state.d} = Medium.setState_ph(p1,1e5);
rho = state.d
m_flow = 1e-4*sqrt(rho*(p1-5e5));
// p1 is the tearing variable, state.d and rho are torn

{state.T, state.a, ..., state.phase} = Medium.setState(p1, 1e5)
// solved by re-computing state := Medium.setState_ph(p1,1e5),
// or just by keeping the last results computed at the previous time
// step and held in some auxiliary variable containing all the function
// outputs

In other words, the state.phase unknown integer is not needed to solve for p1, m_flow, and rho, and there are no further variables that depend on it, so there is no reason why it should be included in the strong component, ending up in a mixed-integer system of equations.

Incidentally, I do not understand why mixed-integer equations do not show up in Test1. As far as I understand, the structure of the problem is exactly the same, the only difference being that in Test2, function setState_ph calls an external C function to fill in the fields of the record. Maybe this detail could be helpful to find a solution to this problem

[Lennart Ochel]

None of the test models produce back end issues anymore. In some cases there are codegen and runtime issues. Hence, the current implementation has to be improved anyway, but it seem to work fine for the particular examples from above.
Francesco, can you please verify if it is working for you? You will need at least r22641.