Convolutional neural networks for predicting morphological and nonlinear optical properties of thin films of quasi-two-dimensional materials

Two-dimensional materials are promising candidates for the creation of flat photonics devices. The main problem of using such materials for applied applications is the complexity of creating films of specified geometric parameters. The films of two-dimensional materials made by exfoliation or chemical deposition methods are usually randomly distributed over a large area and have a large thickness spread. In this paper, we use convolutional neural networks to predict the film thickness of a quasi-two-dimensional material based on optical microscopy data. Hexagonal boron nitride, which is actively used in the creation of flat electronic and optoelectronic devices, was chosen as a test material. Due to the high spatial resolution of microscopy, it is possible to achieve high accuracy in predicting the thicknesses of flat areas of the sample, which allows for rapid characterization of structures. In addition, using the example of the signal of the third optical harmonic, we show the possibility of predicting the magnitude of the nonlinear optical response of the film, which expands the scope of the method.