Iterative peak-fitting of frequency-domain data via deep convolution neural networks

High-throughput material screening for the discovery and design of novel functional materials requires automatized analyses of theoretical and experimental data. Here we study the subject of human-free analyses of one-dimensional spectroscopic data, e.g. in the frequency domain, via employing deep convolution neural network. Specifically, we trained various deep convolution neural network and benchmarked their performance in decomposing one-dimensional noisy data into multiple nonorthogonal peaks in an iterative manner, after which a conventional basin-hopping algorithm was applied to further reduce residual fitting error. Among six different network architectures, a variant of squeeze-and-excitation network (SENet) structure that we first propose in this study shows the best performance. Dependency of training performance with respect to the choice of the loss function is also discussed. We conclude by applying our modified SENet model to experimental photoemission spectra of graphene, MoS2, and WS2 and address its potential applications and limitations.

Automated summary

Automated analysis of one-dimensional spectroscopic data using deep convolution neural networks, particularly a modified SENet structure, showed the best performance in decomposing noisy data into nonorthogonal peaks. The study also discussed the dependency of training performance on the choice of loss function. The modified SENet model was then applied to experimental photoemission spectra of graphene, MoS2, and WS2, revealing its potential applications and limitations.