Optimizer-Free Sizing of OpAmps Leveraging Structural and Functional Properties

Finding appropriate parameters for a circuit via optimization is time-consuming and can fail for certain initial points. In this work, we use machine learning methods to entirely skip the process of manual sizing and tuning device sizes via nominal optimization. A neural network (NN) with a custom loss function is developed to predict size ratios complemented by a gradient boosting algorithm to predict device lengths. Furthermore, with the help of structural and functional analysis methods, we extract canonical design knowledge from random optimal samples of various OpAmps, such that training for a specific circuit only consists of fine-tuning NNs. We will show in two examples that this method can achieve good results, even with small sample sizes.