Learning Forum

[Sara91]

Hello,

I am doing a 3D simulation for textured 10 um silicon. The structure is 10 um silicon with inverted pyramids texture sitting on a glass substrate. My simulation region is 12 um in length (10 um silicon + 1 um above and below for monitors, source, and substrate) with PML BC. The width is 1 um with symmetrical and anti-symmetrical BC. I am starting with 1 um lattice constant but I want to increase it even further. My problem is that the simulation takes 3-4 days for one run with 1 um width and will double when if I tried to increase the width and I am using the auto-non uniform mesh with accuracy = 2.

I am looking for ideas that I can do with the mesh to get good accuracy but with reduced time (i.e. reduce the mesh along 10 um length of the silicon, ....).

P.S. I am doing a broadband simulation (from 300 to 1200 nm) with 300 wavelength points total across the whole range. I am dividing it into two simulations (from 300 nm to 800 nm and from 800 nm to 1200nm) to be able to do a good material fit for each wavelength range independently.

Thanks.

[Taylor Robertson]

Hello Sara Thanks for the question. FDTD is an approximation free solution to Maxwell's equations, which means it is accurate but can become very computationally intensive. That sort of volume is within the range of possibility, but like you said it is hard to iterate over. From my experience it is wise to minimize your computational requirements initially, and focus on accuracy only after you have lots of confidence in your model. Of course with more\faster computers and extra licenses this could be handled very efficiently. FDTD can be scaled up to run on supercomputers if you have access to such a thing.
The mesh size is determined by the shortest wavelength, and material index. The auto-nonuniform mesh will try and create a mesh that allows ~10 mesh cells for all materials using MA=2. The real challenge with your set-up seems to me the 10um Silicon Layer.
I think that the best approach, at least to start may be to bring the bottom PML boundary up so that it below the textured surface. This would significantly reduce the simulation volume and reduce the run time by an order of magnitude if not more. Just make sure that it is far enough away from the pyramids that it will not interact with the resonant fields of the structures maybe lambda/2 as a good rule of thumb.
I don't know exactly what you are trying to understand, but I don't think that you will lose much from this approach. You will be ignoring the presence of the bottom interface which could be important in some respects. Particularly there will be reflections from this surface that will bounce back. Additionally there will be secondary reflections from these and so on and so forth. Fortunately the layer is thick enough ~30* lambda that you likely wont see significant coherent interaction. So this may not be as accurate as modelling the entire system it will be a very a good approximation.
You could use the STACK solver to get an estimate of the power from the Fresnel equations, which would be a few percent I imagine. Ultimately you may want to simulate the entire system, but maybe you can reduce the number of candidates before then.
One way to reduce the number of mesh cells would be to use a coarse mesh override in the thick Si layer. You will want to finely mesh the structure, but could get away with a coarser mesh in the region where the light will propagate more or less freely.
I hope this helps. Please let me know if you have any further questions.

Helpful links.

Best