[PS-1-04] Generative Adversarial Network Applied to Neuroevolution-based Compact Device models
We propose a novel generative adversarial network (GAN) process to optimize the machine learning compact device model (MLCM) based on variable length genetic algorithm (GA), ADAM, and HSPICE. Slightly different from standard GAN, in this work, the generator tries to generate accurate MLCMs while the discriminator generates ill-convergence circuits that will be fed back to the generator during MLCM formation. The generator aims to boost the MLCM quality, and the discriminator is designed to locate the most difficult circuits regarding HSPICE convergence. With GAN, we obtain the optimized MLCM with high fitting accuracy and improved HSPICE convergence.