Estimating the actual parameters of real holographic volume gratings from diffraction efficiency measurements is challenging. The natural formation of the grating provides different phenomena, such as shrinkage, bending of the fringes, or non-homogeneous modulation as a function of the thickness, amongst other issues. This work proposes a deep learning Convolutional Neural Networks (CNNs) and Feedforward Neural Networks (FNNs) hybrid architecture capable of predicting the grating parameters from theoretical and experimental diffraction efficiency patterns. For the training set of this regression problem, Kogelnik’s Coupled Wave Theory simulated data has been employed. Our best model has been trained with an 8000-sized dataset of 80 points of diffraction efficiency patterns simulated from a range of values for the normalized grating wavelengths, index modulations, and optical depths. It achieves test losses under one per cent (mean absolute error) for predicting the normalized grating wavelengths, index modulations and optical depths.
|