*********************************
There is now a CONTENT FREEZE for Mercury while we switch to a new platform. It began on Friday, March 10 at 6pm and will end on Wednesday, March 15 at noon. No new content can be created during this time, but all material in the system as of the beginning of the freeze will be migrated to the new platform, including users and groups. Functionally the new site is identical to the old one. webteam@gatech.edu
*********************************
Title: Design Space Extrapolations and Inverse Design using Machine Learning
Committee:
Dr. Swaminathan, Advisor
Dr. Lim, Chair
Dr. Raychowdhury
Abstract: The objective of the proposed research is to investigate machine learning techniques for power delivery, signal integrity and EM problems. Two broad design strategies have been analyzed. Often one needs to predict the structure behavior outside the range of simulations. This work deals with extrapolation in two domains. (1) We propose HilbertNet for complex-valued causal extrapolation of frequency responses. The proposed architecture accurately predicts the out-of-band frequency response by modelling the temporal correlations between in-band frequency samples using specialized recurrent neural networks. The proposed architecture has been applied to a wide variety of applications including transmission lines, RF filters and power distribution networks. Furthermore, we quantify the uncertainty of our predictions in the extrapolated band using Bayesian inference and approximate it using variational techniques. Along with providing the mean output prediction for out-of-band frequency values, we provide the output variance as a measure of uncertainty as well. (2) We propose Transposed Convolutional Networks to model spatial correlations in the design space. The design space comprises of all the geometrical and material parameters characterizing the response. The convolutional networks can extrapolate the design space in as high as 11 dimensions because of inducing spatial bias into the model. The technique is applied to 5G band RF filters as well as power delivery applications. Furthermore, we also focus on inverse design of electronic systems. The goal in inverse design is to estimate the best set of design space values that generate the response space. We employ invertible neural networks to model the non-linear mapping between the design space and the response space. Using invertible networks, we can find the exact posterior distribution of a high-dimensional mapping design space for a given desired target. The technique has been applied in the areas of high-speed signaling and power delivery.
