Invertible Neural Networks for Inverse Design of CTLE in High-speed Channels

Abstract

Designing CTLE of high-speed channels can be complicated and time consuming. To alleviate this issue, this paper investigates the invertible neural networks (INNs) for inverse design of the CTLE. In this approach, a desired eye height and eye width is given, and the algorithm finds the corresponding peaking frequency and gain value of the CTLE. INN is a special type of neural networks that can be traversed in both forward and reverse directions. An advantage of this network is producing distribution of the input variables based on the desired output. This feature enables the algorithm to provide multiple solutions when a multi-modal distribution is produced. Thus, the user can choose the appropriate solution based on other constraints. A numerical example for inverse design of CTLE of a SerDes channel is provided, which results in moderate accuracy. However, other variations of the example show that the accuracy is case dependent which implies improvements on the algorithm is needed.