Discussion Forum

Hello, I’m trying to add a Model Order Reduction (MOR) study to get a reduction model from a simple unreduced moving spring model. I’ve encountered a problem with the Modal study though. In short: the unreduced model works fine but when I look at the results of the reduced model there is something wrong; the probes which measure the input voltage reflect either a constant voltage or go to a really big (positive or negative) number depending on how I define the “training expression” in the Model Reduction study settings. I will explain my model which I use for the Model Reduction study (Modal) below:

The (2D) model is a simple small beam where a terminal Voltage function is applied to (Electrostatics Physics). The terminal is on all the boundaries of the beam. Under the beam is a plate with constant terminal Voltage, also on all its boundaries. With Solid Mechanics Physics the plate with constant Voltage is completely held in place with a domain fixed constraint node and the beam is only fixed on the left side with a boundary fixed constraint so it's able to move in the y-direction (to and from the plate). The idea is that with a Voltage function through the beam, it will actuate harmonically in the y-direction (perpendicular to the longest length). To enable an electrostatic force on the beam through the Voltage difference between the beam and the fixed plate I introduced air surrounding the two structures so the Charge Conservation (gas) can be applied and handle this interaction. For now, I defined the input Voltage function on the beam as a simple sinus waveform with time as its variable. To make sure the Physics is a Neumann condition (required for the Model Reduction study) I checked the “Use weak constraints” box for the input voltage terminal. I added: “Deforming Domain” as “Moving Mesh” in Definitions for a more accurate model. This time-dependent model works as intended. TL;DR: it’s a simple 2D model where a beam driven by a voltage function moves up/down above a plate and this model is used for a Model Reduction Study.

When I use the model reduction study (Modal) the problem arises. For the study, I used an Eigenvalue study to get the Eigenvalues and use it as the training study. I also used the time-dependent study as the Unreduced model study setting. To get the results of the reduced model I use probes. First, a point/boundary/domain probe to measure the desired output (for now the Voltage of the moving beam to see if it correlates with the input Voltage). In the MOR study, I define the outputs as the expression of these probes. Then I add a global variable probe for each point/boundary/domain probe, these global probes are turned off in the “Results while solving” setting of the unreduced model. Then after I evaluate the MOR study I enter in the expression section of the global probes the expression of the MOR study output. This way the MOR outputs are stored in these global probes which I then use in a second time-dependent study (where all physics is off and only the global probes are used in the “Results While Solving” setting) to get the results. In the results, I will compare the unreduced and reduced probe table graphs which should look a lot alike since both the global and the point/boundary/domain probes measure the input voltage function. This setup for the MOR study and its results is heavily based on this thermal model where they use MOR: https://www.comsol.com/model/thermal-controller-reduced-order-model-56861 TL;DR: I use a setup for MOR based on the model link above, this model uses probes to first measure the desired output in the unreduced model, these probes are then used as outputs in the MOR study, these outputs are used in global variable probes, these global variable probes are used to calculate the results of the reduced model in a second time-dependent study.

The results of the global probes are faulty. The results are either very high/low (-+10^273) or the same as the constant I used as the training expression. This seems to be dependent on what value I use for the training expression (0, 1, -1, the amplitude of the sine input voltage). The results show the same faults but on different training expression values if I use a linear input function instead of a sine. But I also notice that the results sometimes changes while the model and the settings are unchanged, suggesting there is a random part involved with the faulty results. But no matter what I do the results won’t reflect the input (a normal sine wave). I thought that maybe one of you knows what is going on. For your knowledge, I use Comsol Multiphysics 5.5 and the model is 2D.

Thank you for your help. Gyan