We propose an artificial intelligence (AI) modeling method that utilizes an arbitrary voxelized mesh of a small PCB via structure, without relying on parametric geometry. Our AI model, based on a multilayer convolutional neural network (CNN), significantly reduces simulation time for small PCB via structures. By employing advanced AI techniques, we enhance the accuracy and quality of predictions while substantially reducing the size of the neural network model compared to our previous work using a fully connected neural network (FCN). The CNN-based model predicts 2-port scattering parameters with a median Huber error of 0.16%, achieving computational speeds over 13,000 times faster than traditional full-wave field solvers.

In our work, we demonstrate a novel Neural Network (NN) based methodology for generating S-Parameters based on using the meshed and discretized geometry as input to the NN. Our approach would enable a designer to rapidly optimize the performance of vias across an entire substrate. This performance is possible while also achieving a computational time improvement of approximately double that in other NN-based works.

Hardware design for radio frequency devices often demands the use of full-wave electromagnetic (EM) field-solvers to extract scattering parameters (S-parameters) for systems that contain discontinuities throughout their printed circuit boards (PCBs) and substrate packages. Using EM field-solvers not only requires extensive EM knowledge during the setup phase, but also demands considerable computational resources. Previously, machine learning (ML) and neural network (NN) models have been used to solve EM fields; however, they tend to use design parameters as input rather than the structure's geometry. Parametric analysis is beneficial when the solution space is well bounded, but it is more desirable to develop the capability of feeding generic and arbitrary geometry into a machine learning model for rapid S-parameter extraction. This work presents an algorithm for converting a subsection of hardware geometry designed with an electronic design automation (EDA) tool into a dataset suitable for training and validating machine learning models including fully connected NNs and convolutional NNs (CNNs).

Transmission line geometry is one of the most critical aspects of high-speed digital and radio frequency (RF) printed circuit board (PCB) design. While relatively simple equation-based methods exist to estimate transmission line parameters such as characteristic impedance, they do not hold accuracy beyond certain structural limitations such as track width to dielectric thickness ratios. Electromagnetic field solvers have become far more common in recent times and can deliver exceptional accuracy with relatively low computational cost. This paper describes a proof-of-concept neural net which utilizes five input parameters of an uncoated microstrip transmission line and is able to output the per-unit-length equivalent resistance, inductance, conductance, and capacitance RLGC parameters. The goal of a deep-learning based transmission line tool is to enable  accurate microstrip and stripline PCB trace design with computation speeds which allow the engineer to compute hundreds or thousands of iterations in a short period of time with commodity hardware. The final model calculates characteristic impedance with less than ±1.33 Ω error for 100,000 swept samples when the samples are within the middle 90% range of the training data and with a computation speed increase of 14.5 times faster than the benchmark field solver.

Design and analysis for electromagnetic compatibility (EMC) often demands the use of full-wave electromagnetic (EM) field-solvers to extract scattering parameters (S-parameters) for systems that contain discontinuities throughout their printed circuit boards (PCBs) and substrate packages. Using EM field-solvers not only requires extensive EM knowledge during the setup phase, but also demands considerable computational resources. Previously, machine learning (ML) and neural network (NN) models have been used to solve EM fields; however, they tend to use design parameters as input rather than the structure’s geometry. Here, we present a functional convolutional NN which is trained on a single transition via in a typical multi-layered PCB substrate using only a three-dimensional (3D) voxelated grid representing the conductor and dielectric arrangement as input, and provides the 2-port S-Parameter matrix as output. Our NN has better than 1% Huber loss, while its computational time is 6,500 times faster than a conventional field-solver. Our NN-based modeling methodology serves as proof-of-concept for efficient approximates of S-parameters and may be considered as a fast and inexpensive tool compared to the expensive and often slower full-wave EM field-solvers.