Application of Artificial Neural Networks to the Automation of Bandgap Reference Synthesis

Abstract

Bandgap voltage references are present in virtually every analog/mixed-signal system. However, their design still remains a time-consuming procedure that requires extensive designer expertise and validation. In this paper, an automated bandgap synthesis procedure is used to generate a dataset that maps the specifications of the synthesized bandgap reference circuit to their corresponding designer's degrees of freedom. This dataset is then used to train a neural network to predict the choice of the degrees of freedom in order to meet arbitrary circuit specifications specified by the user including variations due to design corners and random mismatch. The automated bandgap synthesis procedure uses precomputed look-up tables rather than invoking a circuit simulator in the loop, which enables generating a large dataset of training examples in short time. The choice of the degrees of freedom predicted by the neural network is then re-fed to the bandgap synthesis procedure to verify the accuracy of the prediction and obtain the complete solution of the synthesized circuit. The results demonstrate that the trained neural network is capable of making successful predictions of good accuracy in a wide multi-dimensional design space.