Designing Energy Efficient Neural Networks According to Device Operation Principles

This article compares the accuracy and efficiency of two neural network architectures for estimating the performance metrics of a modified uni-traveling wave carrier (MUTC) photodetector: a fully connected neural network (FCNN) and a recurrent neural network (RNN)-like architecture designed specifically for the unique properties and operation principles of the device. The RNN-like architecture mimics the current flow and wave propagation inside the MUTC PD, with distinct neural network layers used to pair parameters of each PD layer, inform the network about interfaces between layers, mimic current transport between neighboring layers, and teach the network about the order and bi-directional interaction among the paired inputs. Results show that while the RNN-like architecture exhibits slightly higher learning accuracy than the FCNN, the RNN-like architecture consumes less energy and requires less time for training due to using fewer neurons. This study highlights the potential of physics-inspired neural networks in solving problems in electromagnetics and photonics.