ABSTRACT
With the advancement of integrated circuit manufacturing technology,
more and more aspects must be considered during the electrical
characterization of circuits in order to solve challenges such as process
variability effect. This increases the characterization time due
to traditional techniques based on exhaustive electrical simulations.
The adoption of machine learning techniques already helps digital
design at many levels of abstraction. Thus, the main objective of
this research is to evaluate machine learning regression algorithms
as an alternative to exhaustive electrical simulation in the cell characterization
project. In this step, multiple linear regression, support
vector regression, decision trees and random forest algorithms were
considered. This work presents the results of NAND2 and NOT
gates using bulk CMOS technology. Specifically, the energy values
and the propagation times of this circuit will be predicted separately.
A comparative analysis, together with the inference time,
is made for each dependent variable between the models, in order
to understand which is the best regression model for the task. The
algorithm with the lowest cost function and shortest inference time
proved to be the decision tree for all predicted variables in both
gates.