Data Driven Surrogate Models for Faster SPICE Simulation of Power Supply Circuits

Abstract

The stiffness of power supply circuits with large power distribution networks makes simulation through the industry standard Simulation Program with IC Emphasis (SPICE) often non-convergent or prohibitively expensive. The existing solution of piecewise linear (PWL) simulation addresses these issues with reasonable accuracy, but lacks practicality. This research implements a system to train surrogate models for an n-type MOSFET that can replace the nMOS device in SPICE simulation to improve performance while maintaining accuracy, regardless of the larger circuit context. We explore a variety of surrogate modeling and adaptive sampling techniques on low-dimensional functions, showing that adaptive sampling improves surrogate prediction accuracy compared to space-filling sampling. We implement a deep feed-forward artificial neural network (ANN) surrogate for the MOSFET, but the current implementation fails to achieve sufficient accuracy to be useful in SPICE simulation. Future work might explore hyperparameter search and alternative neural architectures or adaptive sampling approaches to improve accuracy.