E-field simulation of high-frequency (RF) circuits

We have the following geometry of a stripline-via-stripline transition.

In blue are the sampled points in the air or dielectricum region and red are all sampled points that belongs to the stripline or via.

The eigenmodes are to be specified at the input port (left) of the geometry. The PINN should then determine the E-field using Maxwell’s equation.

We are currently using a fully connected architecture for the neural network. We have defined the boundary conditions as follows:

PEC for all boundaries of the trace geometry with the exception of the input port on the left and the output port on the right. PML for all outer boundaries of the air domain. All points within the trace geometry should have E-field = 0 in all three dimensions. The input port was defined with a set of eigenmodes from a .csv file. The output port was defined as PML.

Unfortunately, the resulting solution is not correct at all and it is unclear to me where the problems could lie. Does anyone already have experience with such problems?

Many thanks