Design Space Extrapolation for Power Delivery Networks using a Transposed Convolutional Net

Abstract

The geometrical and material properties of distributed electromagnetic structures comprise the design space. This space characterizes the structure’s frequency response in complex domain. In this paper, we propose a machine learning framework for predicting frequency response of a power delivery network as a function of its extrapolated multidimensional geometrical and material parameters. The proposed approach comprises of an ensemble of architectures: (1) Fully Connected Upsampler for latent code generation (2) Convolutional Decoder to learn the frequency response from the latent code. The 14D design space is converted to a Lth dimensional code which entails the frequency response information. With the proposed architecture, a root mean squared error of 0.004 ohms is achieved when compared to the true value. We focus on extrapolation of design space parameters while training on in-band values. We also illustrate how frequency poles move with varying design space exploiting parameter sensitivity in different frequency bands.