The Investigation of Deep Generative Models for Statistical BSIM Parameter Generation

Generating compact model parameters is a commonly used method for producing statistical compact models in large-scale integrated circuit simulation. This article focuses on exploring the capabilities of the advanced generative models in compact model parameter generation, so that the generated parameters should not only follow the distributions of the original parameters but also maintain the interparameter correlations simultaneously. In this article, a variety of generative models, such as vector quantized-variational autoencoder (VQ-VAE), masked autoencoder for distribution estimation (MADE), adversarial autoencoder (AAE), VAE-GAN, and diffusion models, are theoretically analyzed and experimentally compared for the BSIM parameter generation. The results show that the VAE-GAN stands out for its excellent performance, especially in capturing the tail distribution, which is vitally important as the tail determines the circuit yield. The parameter generation capability of VAE-GAN is further verified when comparing with other research findings via experimental data. We also verified the reliability of VAE-GAN in generating compact model parameters at the circuit level. This article guides deep-learning-based parameter generation tasks.