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 theoretical experimental data. Here we study subject human-free one-dimensional spectroscopic data, {\it e.g.} in frequency domain, via employing deep convolution neural network. Specifically, trained various network benchmarked their performance decomposing noisy data into multiple nonorthogonal peaks an iterative manner, after which a conventional basin-hopping algorithm was applied to further reduce residual fitting error. Among six different architectures, variant "Squeeze-and-excitation" (SENet) structure that first propose this shows best performance. Dependency training with respect choice loss function is also discussed. We conclude by applying our modified SENet model photoemission spectra graphene, MoS$_2$, WS$_2$ address its potential applications limitations.