Machine learning assisted GaAsN circular polarimeter

TL;DR

This paper presents a machine learning approach that accurately correlates electrical signals from a GaAsN circular polarimeter with light's polarization properties, improving device performance even with limited training data.

Contribution

It introduces a novel two-stage machine learning algorithm, combining logistic regression and neural networks, to enhance the accuracy of polarization detection in GaAsN devices.

Findings

Achieves over 97% accuracy in predicting polarization states.

Correctly classifies polarization with less than 1.5% error.

Effective even with small experimental datasets.

Abstract

We demonstrate the application of a two stage machine learning algorithm that enables to correlate the electrical signals from a GaAs$_x$N$_{1-x}$ circular polarimeter with the intensity, degree of circular polarization and handedness of an incident light beam. Specifically, we employ a multimodal logistic regression to discriminate the handedness of light and a 6-layer neural network to establish the relationship between the input voltages, the intensity and degree of circular polarization. We have developed a particular neural network training strategy that substantially improves the accuracy of the device. The algorithm was trained and tested on theoretically generated photoconductivity and on photoluminescence experimental results. Even for a small training experimental dataset (70 instances), it is shown that the proposed algorithm correctly predicts linear, right and left circularly polarized light misclassifying less than $1.5\%$ of the cases and attains an accuracy larger than $97\%$ in the vast majority of the predictions ($92\%$) for intensity and degree of circular polarization. These numbers are significantly improved for the larger theoretically generated datasets (4851 instances). The algorithm is versatile enough that it can be easily adjusted to other device configurations where a map needs to be established between the input parameters and the device response. Training and testing data files as well as the algorithm are provided as supplementary material.