Solutions to intermediate evaluation by CVODE are leaking to algebraic loop solutions

[rruusu]

Description

The initial guesses for each algebraic loop evaluation are not based on the end of the last solver step. Instead, the initial guess is based on whatever result was last obtained by the previous RHS evaluation of the ODEs, which might be related to a trial step in Jacobian evaluation using difference quotient steps.

This results in both unnecessary noise to inputs in algebraic loops, as well as possible extra iterations.

Version and OS

• Version: OMSimulator v2.1.2.post5-gce342b6-mingw-notlm-debug
• OS: Windows 11, 64-bit

[rruusu]

I will push a solution for this to my fork shortly.

[rruusu]

AlgebraicLoopTest_zero.zip

Here is a small model to demonstrate this phenomenon on a rotational mechanics model of two connected shafts as separate FMUs. There is an algebraic loop between the torque and acceleration at the intersection of these FMUs.

The script sets all damping values to zero and provides no exiting torques to the ends of the shaft. As the initial state is at rest, the entire simulation result should only consist of zero speeds.

However, when simulated with the current head of maintenance/v2.1, we get these accelerations at the ends of the second FMU:

And this torque at the interface between the components:

This is caused by the algebraic loop solution accepting non-zero inputs generated by the CVODE solver as a part of the numerical Jacobian estimation process, in which states are slightly perturbed.

Due to the zero damping, these exitations from these small perturbations accumulate and are made easily visible.

With the solution at #1525, we get this, as expected:

It forces each algebraic loop solution to use the final result of the last solver step as its initial guess.