A Differential Evolution based Methodology for Parameter Extraction of Behavioral Models of Electronic Components

Abstract:
Behavioral models of electronic components are crucial for system simulation, as they are quick to simulate and yet provide reliable information on the behavior of the original circuit. Parameter extraction of such models, i.e. calibrating the model to match the experimental characteristics of the device, is tedious work, as the number of such parameters can add up to tens of them. There are attempts in the literature to solve this problem with the help of optimization algorithms. However, when put to practice, new challenges arise due to the large number of devices to be calibrated, time restrictions and the wide variety of behavioral models. These challenges require novel techniques that ensure generality, speed and scalability. We address these challenges by proposing a fully automated flow, which includes the following novel features: an evolutionary algorithm, a smart sampling technique for reducing redundancy in the reference data, and a method for making use of the knowledge acquired in previous parameter extraction tasks. We tested the flow with a set of more than 200 Si-diodes and IGBT behavioral models with more than 50 parameters and 30 response curves to be calibrated. The results show that full automation of parameter extraction is possible, i.e. no human intervention is needed. Hundreds of Si-diodes and IGBT behavioral models are calibrated within 48 hours as compared to 1.5 years of manual work.