AI-Assisted Raman Spectral Peak Label Assignment

Abstract

Raman spectroscopy is a robust technique that enables fingerprinting of organic and inorganic materials. By measuring the vibrational modes of molecules, it is possible to extract information about the chemical structure, crystallinity, and molecular interactions of a given sample. For simple molecules or crystal materials, interpretability can be achieved by considering the selection rules and characteristic frequencies of know vibration groups. In these cases, a skilled user can assign each peak to a given functional group or chemical bond and relate changes to peak position, width, and/or intensity to material properties. The problem, however, lies in complex or unknown materials, when peak assignment requires either manual searching the literature or computationally intensive simulations. Although many AI-assisted classification models have been demonstrated for spectroscopy applications, most have focused on identifying molecules by comparing the whole spectra with databases of known molecules. In this work we explore how AI techniques can assist in the interpretation of spectra by predicting which chemical bonds are most likely to be related to a specific Raman peak. We are developing Raman peak label assignment models using AI methods and evaluating their performance. The dataset consists of 1000 Raman spectra and related crystallographic information files (CIF files). The preliminary analysis indicates the dataset has more than two hundred chemical bond types, with a model showing a good correlation (above 80% of accuracy).