Bidirectional Deep Learning of Polarization Transfer in Liquid Crystals with Application to Quantum State Preparation

Abstract

Accurate control of light polarization represents a core building block in polarization metrology, imaging, and optical and quantum communications. Voltage-controlled liquid crystals offer an efficient way of polarization transformation. However, common twisted nematic liquid crystals are notorious for lacking an accurate theoretical model linking control voltages and output polarization. An inverse model, which would predict control voltages required to prepare a target polarization, is even more challenging. Here we report both the direct and inverse models based on deep neural networks, radial basis functions, and linear interpolation. We present an inverse--direct compound model solving the problem of control voltages ambiguity. We demonstrate an order of magnitude improvement in accuracy using deep learning compared to the radial basis function method and 2 orders of magnitude improvement compared to the linear interpolation. Errors of the deep neural network model also decrease faster than the other methods with an increasing number of training data. The best direct and inverse models reach average infidelities of $4\ifmmode\times\else\texttimes\fi{}{10}^{\ensuremath{-}4}$ and $2\ifmmode\times\else\texttimes\fi{}{10}^{\ensuremath{-}4}$, respectively, which are accuracy levels not reported yet. Furthermore, we demonstrate local and remote preparation of an arbitrary single-photon polarization state using deep learning models. The results will impact the application of twisted-nematic liquid crystals, increasing their control accuracy across the board. The presented bidirectional learning can be used for optimal classical control of complex photonic devices and quantum circuits beyond interpolation.