Deep learning applied to peak fitting of spectroscopic data in frequency domain

초록

A data-driven study of material properties and functional materials design based on it requires high-throughput and comparative analyses of the results of experimental spectroscopy with those from first-principles electronic structure calculations. Hence, an efficient machine-learning-based computational tool to extract electronic structure information from experimental data without human intervention is in high demand. Here, we test the capability of deep neural network models to fit photoemission spectroscopy (PES) data in the frequency domain with unknown PES peak positions, numbers, and widths. A one-dimensional convolution neural network (CNN) was employed in combination with fully connected layers (FCL), and the trained model was applied to photoemission spectra for the sulfur 2p states in poly(3-hexylthiophene) (P3HT) molecules and oxygen 1s states in indium tin oxide (ITO). We conclude by further discussing potential ways to improve the performance of the model. © 2020 The Korean Physical Society. All rights reserved.

키워드