Circuit convergence study using machine learning compact models

Abstract

Machine learning (ML) compact device models (CM) have emerged as an alternative to physics-based CMs. ML CMs can find a mathematical model close to the device characteristics without much prior knowledge, which saves the time of model formation. Additionally, versatile capabilities such as process-awareness, model merging, and fitting new technologies, promote the usage of ML CMs. While ML CMs draw great attention in CAD, their convergence in SPICE has not been carefully studied. Here different activation functions are used to create ML CMs, and then the circuit convergence is tested. We found that inverse square root unit (ISRU) activation has the best convergence. Besides, gate-to-source and gate-to-drain capacitance is founded to benefit the convergence in transient analysis. The circuit convergence rate is 100% for ISRU, sigmoid, and tanh when the capacitor is present. On the other hand, ISRU significantly outperforms other activation functions in DC sweep, achieving 81% convergence. If quasi-static transient analysis is employed to replace DC sweep, 100% convergence is achieved by ISRU. Due to its superior convergence, ISRU is the most promising for future ML CMs in SPICE.