A Pseudo-supervised Machine Learning Approach to Broadband LTI Macro-Modeling

Neural networks have been popularized by their ability to learn complex, non-linear mappings between features and output spaces. They have been used for learning the mapping between geometry and network parameters for various electrical structures such as interconnects. In this work, a novel neural network architecture is applied to black-box identification problems in which poles and residues of a dynamical system are the quantities to be extracted from frequency domain network parameters. Once poles and residues are extracted, time-domain simulation can be performed using well-establish time-domain simulation techniques.