Application of Deep Convolutional Neural Network for X-ray Reflectivity Analysis

Abstract

X-ray reflectivity (XRR) is one of the powerful techniques that provide structural information on thin film. The development of XRR measurement techniques has led to the need for fast XRR analysis techniques. In this thesis, we introduce the possibility of a new x-ray reflectivity analysis method by applying a deep convolutional neural network to XRR data. With a slab-based sample model, an analysis of complex multilayer samples have been performed. Our convolutional neural network resulted 4~10% accuracy for simple samples with respect to target XRR curve. For a complex sample such as a multi-layer sample, the model achieves a lower accuracy with a relative error of 18.7%, but it is powerful enough to provide an initial parameter of sample for fitting. Nevertheless, the analysis takes 0.8 seconds per single XRR curve, which is much faster than conventional methods. The value of this work lies in that deep learning algorithms can be applied to physical data analysis. Since this work has much room for improvement as it is an initial trial, we expect it to be improved into a meaningful alternative to future XRR analysis methods.