Fast Topology Selection for Analog Amplifier Circuits Using On-The-Fly Cascaded Neural Networks

Abstract

In this paper, a machine learning-based approach for the automation of topology selection of integrated analog amplifier circuits is presented. A dataset of 480,000 circuits for 30 different amplifier topologies is generated for the prediction algorithm based on a precomputed lookup tables (LUTs) approach. A first approach based on neural networks is presented where the required specifications act as inputs to the networks, and the output of the network is the suitable topology for such a set of specifications. A modified cascaded neural network approach is examined to reduce the training time of the network while maintaining the prediction accuracy. Using the cascaded neural network approach, the network is trained in only one minute on a standard computer, and a 90.8% prediction accuracy is achieved. This allows on-the-fly changes in the input specifications, and consequently the neural network, to enable examining different design scenarios.