Automatic Optimal Design Method for Circuit Sizing Based on CNN Surrogate Model Assisted Differential Evolution Algorithm

The physical characteristics of analog ICs are intricate, leading to prolonged design and development cycles. Currently, the automatic optimization of analog integrated circuit size heavily relies on simulations, which unfortunately results in a significant waste of simulation resources during the optimization process. To address this issue, this paper introduces a novel optimization framework named "MODE-CNN", which integrates Multi-Objective Differential Evolution (MODE) algorithm and Convolutional Neural Network (CNN) surrogate models. The MODE-CNN framework is designed to reduce unnecessary consumption of simulation resources while maintaining high optimization accuracy. Within the MODE-CNN framework, we first enhance the predictive capabilities of the surrogate models using Latin Hypercube Sampling and Quantile Transformation. Subsequently, we employ CNN surrogate models to evaluate circuit performance during the optimization process, effectively screening out promising designs and substantially reducing the number of simulations required. Moreover, we propose an improved fast non-dominated sorting method that further enhances the global optimization performance of the MODE algorithm. Through testing on three circuit design cases and comparing MODE-CNN with existing mainstream methods and expert manual designs, we find that MODE-CNN not only significantly reduces the number of simulations (by up to 63%), but also surpasses traditional methods in optimization outcomes, with performance improvements of up to 95.09% for TC and 58.75% for LS. The experimental results validate the effectiveness and advancement of the MODE-CNN framework in the optimization of analog integrated circuits.