上海科技大学知识管理系统

Analog Circuits Design typically requires a lot of handwritten formula derivation to work out the performance metrics of circuits. With the help of SPICE tools, we can predict the metics easily. But the process of running SPICE simulation tools is still tedious when the scale of circuits parameters gets large. In the article, we combine Graph Neural Network (GNN) with Equation Learner Network (EQL) as pretrained model in small range of design parameters. We show that datasets built with the gm/ID method can capture more efficient points in the design space than the W/L random sampling method. We can see that the pretrained GNN-EQL model outperforms other pretrained model when the range of design parameters increases for three different amplifiers. We further see that few shots of our model can have 20X sample reduction compared with training a model covering a wide range of design parameters from scratch.