Machine-learning based analysis of time sequences for multiplexed microresonator sensor

Abstract

This paper discusses an application of machine-learning solution for processing of the dynamical sensing responses collected with a multiplexed microresonator detector. Performance of a long short-term memory network (LSTM) out of bidirectional and dropout layers is analyzed on example of the experimental data collected for a temporal gradient of the local refractive index. We experimentally demonstrate the possibility for analyte parameters prediction with accuracy of > 99% based on a set of complex non-linear highly specifc time sequences of the intensities radiated by the microcavities which is obtained within a timescale 4 times shorter than required to reach the steady state. Optimization possibilities in terms of the number of microresonator signals to consider for the LSTM network training along with the complexity of its architecture are analyzed. Keywords: optical microresonator, sensing, machine learning, whispering gallery mode, multiplexing