Simulation results
First run
Gautam ran the COMSOL model on sandbox1 since we were trying to run it before I had a home directory there to run from my login. Since that first run, Larry set me up on sandbox1 so I was able to run a few more times with some tweaked model parameters.
Here are the results from the first run, which uses the nominal aLIGO test mass parameters:
I don't think I was expecting such a regular pattern here. Are the modes really that closely and regularly spaced? This might also be indicating some problem with my meshing, I could imagine a few scenarios. To go from this chart to something that looks like the suspension thermal noise plot, I'll have to scale this by the frequency-dependent loss angle and a strain sensitivity TF probably? I need to remind myself for another few minutes on this, maybe chat about it.
Relevant properties of the file that might need to be tweaked:
- F_load
- The applied force may have been resulting in large displacements of the test mass (and thus large curvature in the fibers). Since the model contains geometric nonlinearities (I think it has to, no?), large displacements may not be able to be linearly extrapolated back to the small displacements we expect. Since I am running a frequency domain study, COMSOL expects a load force, which means I am specifying a force rather than a displacement. In future runs, I wanted to reduce the displacement by scaling down the force... I'll discuss my first attempt at this later.
- skin_depth
- Gabriele mentioned that when he models surface losses for wafers he does not explicitly mesh a boundary layer, but rather asks COMSOL for a boundary integral (so just the 2D integral on the outer surface of the wafer), and then presumably makes some assumption about the thickness of this layer to go from J/m to J if need be. In contrast, my suspension thermal noise model defines the 'surface' of the fiber as the 15um outer layer, and I explicitly mesh that surface volume and do a 3D (volume) integral over the surface domain.
- Using a 2D integral would significantly simplify the model and probably run faster and result in fewer bugs when the scale of the geometry changes the necessary fineness of the mesh. However, I am a bit uncertain about whether this is what we want. At least in my understanding, we are interested in asking about what happens when the surface layer depth becomes comparable to the radius of the fiber. Since the 2D integral would always represent an infinitesimal volume, and the geometry of an extended (deep) surface layer would differ significantly from that of an infinitesimal shell, so I am concerned that the strain energy density would vary significantly with the radial postition in the fiber. Nonetheless, even the mesh I am using now has only a few (2-4) mesh elements along the radial direction, so I'm not sure I'm doing much better even with the additional computation time. Therefore, in future runs I'd like to try Gabriele's suggestion.
- Frequencies
- I was using 50Hz spacing on the frequencies from 60Hz to 5kHz. I haven't been very systematic about this, but I'm getting some convergence issues going to lower frequencies--maybe by taking the suggestion for eliminating the skin_depth I can make a finer mesh and go to lower frequencies. This frequency sweep is too coarse to find well-defined resonances, but they are suggested in the plot below.
More recent runs
I wanted to automate this loop over mass parameters in a matlab script, so I set up the Livelink handshake so Matlab would send the model to sandbox1 for solving, then MATLAB on my machine would work with the solved model to extract results. I realized later during running that this might not be optimal, since it will require me to keep the connection to the remote COMSOL server during the entire run, which is A. annoying and B. risky because these might take many hours to run and the connection can easily be severed, wrecking the job. I don't really have a workaround for this, so my current plan is to continue logging on to the remote server and running individual COMSOL files. As I'll discuss, this is probably necessary for now anyway since I ran in to some problems running these models, and might want to tweak the models on my local GUI before running them over a large (frequency) parameter space on sandbox1.
aLIGO parameters, but decrease F_load
Based on the plots generated from the first run, I estimated that the TM displacement was comparable to the TM thickness; just as a first pass I figured 'small displacement' could be easily defined as 'displacements smaller than the thickness of the fiber' or maybe 'smaller than the thickness of the surface layer'. If the displacement is directly proportional to the force, this meant I wanted to scale the previous F_load by 1e-5, which I did. I wanted to just see what kinds of displacements I would get, so I asked to run the model only at 300Hz. After more than 45 min of running, the server threw an error that it couldn't find my STEP file for the horns, which were in a local directory but not in the remote (server's) directory.
I thought this was an odd error for two reasons. First, I had originally not even had the STEP in my local directory (it only needed to load once and then could be used many times for a given model in the GUI), and I was getting an error within a couple minutes of starting the job that it could not find the file. Adding the file to my local directory seemed to solve this problem, as the model was running for much longer. The error I got after it finally crashed was not the same error as before, but was still an error in loading the file, which makes me a bit confused about where it is actually looking for this file or when exactly the loading process happens. To solve this, I have copied the STEP to the remote directory and will run again with the import pointing to that remote file, which I suspect will solve the problem.
aLIGO with the mass scaled by 0.1, also decrease F_load by 1e-5*0.1
Since the first run seemed to mostly work up to the outstanding questions mentioned above, I decided to also run the model in batch mode directly from the sandbox1 command line, just as Gautam had. That is, I ssh on to sandbox1 and run the command
$ /usr/local/comsol53/multiphysics/bin/glnxa64/comsol batch -inputfile [inputname].mph -outputfile [outputname].mph
For this run, I scaled the mass of the model by 0.1, so we now have a 4kg TM. Most lengths of the model then scale by 0.1^(1/3), but those related to the radius of the fibers (and the bend point of the fibers) scale by 0.1^(1/2) because we want to maintain a constaint stress in the fiber. The main length of the fiber remained the same so the modes would be in about the same place. I document these more thoroughly in the matlab script, which I should upload to the git. I scale the force by 1e-5 as above (same back of the envelope calculation), with an additional factor of 0.1 to account for the lower mass.
The model ran in to some convergence issues after a few frequencies. I could solve this by changing the relative tolerance, but I think I will most likely instead pursue Gabriele's suggestion above and try to refine the mesh to improve the convergence. The solver nodes of the COMSOL models are still a bit mysterious to me--I don't really know much about when different methods or measures of convergence are appropriate. Probably playing with these could both improve the accuracy and efficiency of the model... but it seems like a hefty undertaking. Might be worht the long term payoff though.
The first few frequencies did converge, but as I was extracting the resuls ancha, the COMSOL license server, went down so I'll have to extract them later.
NOTE TO SELF: WHAT'S WITH THE INLINE IMAGE QUALITY? DON't USE JPG !!!! |