Closed JakobEliasWagner closed 2 years ago
I am not sure why you mean by "varying the wavenumber to get a broader result" and "taken from a "static" approximation".
In regards to the pde $\nabla^2 E_z(x,y)+\varepsilon_r(x,y)k_0^2E_z=0$ (6). In figure 1 the inputs to the network are only x and y. Is that the case for the training as well, or have you tried varying $k_0$, e.g. using it as an input to the network as well? For me using more inputs makes the network very unstable and makes achieving convergence very hard. Thus i would like to know if you tried using other parameters as inputs :) I've you have tried that and got convergence, there would be a problem with my code :P
No. I didn't try.
Hey @lululxvi, i've read your paper "Physics-informed neural networks for inverse problems in nano-optics and metamaterials" and liked the results you presented in it. As i am currently working on a similar question (helmholtz), i encounter the problem of varying the wavenumber to get a broader result. Is that something you have tried? The results presented look like they were taken from a "static" approximation. Thank you, Jakob