A Diffusion-Model-Based Methodology for Virtual Silicon Data Generation

Silicon data allow designers to enhance the chip performance by leveraging machine learning techniques. By gaining a deeper understanding of the distributions of interested features within a wafer, designers can predict chip behaviors more accurately. However, real silicon data may not always be available. This work presents a methodology for generating high-quality synthetic silicon data and verifies its effectiveness through several metrics. Silicon features obtained by chip probing (CP) and wafer acceptance test (WAT) are combined to create more comprehensive data, enabling to conduct design-technology co-optimization (DTCO). Unlike the generative adversarial network (GAN) based methodology used in prior work, this work utilizes a diffusion model to generate synthetic silicon data. The Jensen-Shannon (JS) divergence similarity and Frechet Inception Distance (FID) are used to evaluate the distribution and to quantify the quality of synthetic data, respectively. Experimental results demonstrate that the diffusion model is able to extract the multi-feature silicon data distribution more accurately, with an average JS divergence similarity of 0.987 and an FID of 6.28. This methodology enables to generate a substantial volume of silicon samples for extensive silicon data analysis and DTCO acceleration.