Modelling Integrated Circuits Behavior using an Active Learning Approach based on Gaussian Process Regression

Modelling the integrated circuits performance dependencies on design parameters using machine learning regressions is becoming widely used in various applications like circuit design and optimization or verification. For constructing such models, a certain amount of input-output sample pairs needs to be acquired through circuit simulations. Depending on the circuit complexity this process can become very costly in what regards both time and licensing expenses. In this paper, we propose a sampling scheme for minimizing the number of samples needed to create accurate and reliable regression models by using an active learning approach. We explore the possibility of achieving this using Gaussian Process regression within an active learning scheme based on the particular regression model uncertainty. We validate the concept on synthetic functions used as a placeholder for circuit behavior and also for a simulated LDO circuit.