Machine learning concept in de-spiking process for nuclear resonant vibrational spectra – Automation using no external parameter

Abstract

Nuclear resonant vibrational spectroscopy (NRVS) is a relatively new spectroscopic method which measures site-specific vibrational information and is useful in many research areas. It has an almost zero background and is suitable for measuring weak signals but needs a lot of scans to complete one real spectrum. Due to various reasons, some NRVS scans have occasional spike(s), which can introduce fake peak(s) when the averaged spectrum is converted into partial vibrational density of state (PVDOS) and can mislead the deduction of the corresponding structural information from it. For better use of the NRVS spectra with occasional spikes, people have to identify and smooth the sporadic spiky points while leaving all other points untouched. In this publication, we used the concept of machine learning and created a fully automated procedure for screening and smoothing the occasional spiky points in NRVS spectra. The procedure uses the statistical information obtained from the particular NRVS scan to be processed itself and needs neither an external parameter nor the information from other NRVS scans. A corresponding R subroutine code is also presented to batch process large numbers of measured NRVS scans. This work is the first attempt toward organizing an automatic de-spiking process for NRVS scans without using an external parameter.