Resolved periods different but identical seismograms

[lucela]

Hi,

I am trying to do regional simulations using one chunk of 30deg x 30deg. I am testing for different 2 periods by changing the number of elements. The simulations work fine. After meshing in output_mesher.txt I have 2 different periods resolved and DT but my seismograms look very similar, like the period did not change.

Simulation 1 has 64 MPI processes, 64 elements along xi and eta in each chunk, 8 slices along xi and eta in each chunk. There is one chunk of 30deg x 30deg.

Verification of mesh parameters : Minimum period resolved = 25.3712997 (s) Minimum period resolved (empirical) = 26.4444447 (s) Maximum suggested time step = 8.60000029E-02 (s)

for DT : 0.118749999 (s) Max stability for wave velocities = 0.758804023

Simulation 2 has 64 MPI processes, 128 elements along xi and eta in each chunk, 8 slices along xi and eta in each chunk. There is one chunk of 30deg x 30deg.

Verification of mesh parameters : Minimum period resolved = 17.0147381 (s) Minimum period resolved (empirical) = 11.3333330 (s) Maximum suggested time step = 8.60000029E-02 (s)

for DT : 9.50944871E-02 (s) Max stability for wave velocities = 0.607735932

Should these difference in resolution make a difference on the seismograms ?

Thanks a lot !

[rmodrak]

Hi @lucela,

What is your source time function? To begin to see numerical artifacts, you may need to either increase the frequency content of your source, or else coarsen your mesh even further.

For what it's worth, here are some simple 2D numerical results showing roughly what to expect as far as mesh convergence: Numerical Experiments SPECFEM2D.pdf

In comparison to these simple 2D experiments, SPECFEM3D_GLOBE employs much more complicated meshing. So if you notice any interesting or unexpected behavior, please respond further. This is an issue many of us are very interested in. (At least I am very much currently!)

Best regards, Ryan

[danielpeter]

hi,

the mesh resolution only makes a difference if you use a 3D crustal & mantle model (especially moho topography) with a source that also produces higher frequency energy. in your simulations, the source half duration is likely above the minimum period of the coarse mesh. thus, there is little difference to the higher mesh resolution (a part from a slightly better crustal mesh resolution which will affect surface waves).

in case you use a 1D model without topography you should obtain identical seismograms for a long-period source.

[lucela]

Hi Daniel and Ryan,

Thanks a lot. That might be it : I am simulating a Mw 8.1 (Mo = 1.6E28) event, I set the half duration according to the moment, so it's 26,5s. I should set my half duration do zero then.

I didn't change anything regarding the source time function, according to the manual as I set a non zero halfdur it should be a gaussian, but when I plot it after the simulation it's a step function.

[danielpeter]

ok, that should explain it.

for CMT sources, the source time function is a quasi-Heaviside. so, instead of a step, it has a "smoothed" step with a half-duration such that the corresponding moment-rate function is a Gaussian with a width as explained in the manual. notice the difference between the source-time function and its derivative, the moment-rate function. the file plot_source_time_function.txt shows the source-time function, not the rate function, which is the (quasi) Heaviside.

setting a zero half duration will internally make this "smoothed" step very narrow and steep, such that it is almost a Heaviside with "almost" a Delta function as rate function. you will have to convolve the seismograms with a source-time function of your choice to filter out the high-frequency oscillations due to the numerical mesh stiffness. again, these numerical Green's functions are only valid down to a minimum period determined by your mesh resolution (a minimum estimation is given in the output_mesher.txt file).