Abstract
Electromagnetic (EM) extraction of scattering parameters (S-parameters) from printedcircuit boards (PCBs), multi-chip modules (MCMs) and packages is a critical step
in the design of the latest technological semiconductor innovations. This extraction
step is both time and computationally intensive, but modern advances in the field of
machine learning (ML) and artificial intelligence (AI) provide the tools available to
improve these processes. Other work has been done in the area of neural-network
accelerated S-parameter prediction, but no examples have been found whereby pure
meshed geometry is used as the input to the neural network – rather parameterized
geometry is used which cannot be generalized to any structure. While parameterized
geometry is a valid use-case for limited neural network applications, this work pro-
poses a workflow whereby voxelated structures represent the dielectric and conductor
materials in a PCB via structure which is passed directly into a neural network field
solver. This work will first present an introductory formulation of a fully connected
neural network which is capable of predicting resistance, inductance, conductance
and capacitance parameters (RLGC-parameters) of a microstrip transmission line suit-
able as a replacement for a 2-D field solver. Next, we offer a unique algorithm using
ray tracing which converts a PCB via into a voxel mesh and creates data suitable for
input to a neural network capable of predicting S-parameters. Finally, we showcase
two variants of deep neural networks which take arbitrary geometry as input, in the
form of a voxel mesh, and predict 2-port S-parameters with better than 1% error and
more than 13 000 times faster than a conventional electromagnetic field solver.