Analysis of Raman spectra using denoising autoencoder

Abstract

In many experiments, several denoising techniques are generally used to improve signal to noise ratio (SNR), such as averaging, filtering, smoothing and FFT/wavelet based algorithms. In particular, many studies have recently been conducted to apply learning-based methods such as denoising autoencoder (DAE) to denoising in experiments. Raman spectroscopy is a major experimental method that can study the internal structure and properties of molecules using Raman scattering. Especially, TERS (Tip-enhanced Raman spectroscopy) uses sharp tips to make reaction only near a point, enabling more accurate Raman signal analysis. However, in many experimental condition, SNR is often bad because Raman scattering occurs weakly. In order to increase SNR in Raman experiments, data is averaged through experimental iterations and analysis is carried out using Savitzky-Golay (S-G) smoothing. In the case of TERS, however, the repetition of long experiments is not good for experiment stability or cost issues because it makes the tip unstable. There is also a problem that S-G smoothing could erase important information corresponding to the weak signal of the Raman spectrum. In this paper, to solve the above problems, the two methods of applying DAE to the TERS experiment are presented using bulk water example. First, we made DAE learn data that is difficult to conduct S-G smoothing due to low SNR and confirmed that it can restore signals corresponding to ground truth without loss of information. Second, we made DAE learn data before and after averaging and confirmed that it could restore the averaging data with only a small portion of the data before averaging. This can increase the life of the tip and stability of the experiment by reducing the repetition of the experiment. This experiment has been applied to a well-known data sets such as bulk water signals, but it is expected that the same methodology can be applied to new experimental data to further enhance the quality of the TERS experiment.