Learning of Multi-Dimensional Analog Circuits Through Generative Adversarial Network (GAN)

Analog circuits are strictly designed under operational, functional and technology constraints. Together, these bounds create a sparse multi-dimensional design optimization space with the scarcity of labeled analog training data making supervised learning methods ineffective. Accurate approximation of multi-target analog circuits, therefore, requires generation of labeled data around dominant bias and with relevant variance. With such an approach, we explore state-of-the-art semi-supervised, generative adversarial network (GAN) towards analog performance modeling. We report on various multi-target analog circuit classification experiments and demonstrate stable GAN performance achieving 2-5% higher accuracy and utilizing only 10% fully simulated manually annotated labeled data against supervised learning methods.