Raman Spectrum Prediction from Crystallographic Information with Graph Neural Networks

Raman spectroscopy is a powerful technique for analyzing the vibrational modes of materials and obtain information about their chemical structure, crystallinity, and probing their interactions with the chemical environment. Typically, spectral peak positions, line widths and scattering amplitudes are analyzed and correlated with their microscopic origins withing a material’s structure. A challenge, however, arises in case an unknown material is investigated. Correlating the measured Raman spectrum with a specific molecular structure, requires either ab-initio molecular simulations or literature search, which is time consuming and computationally costly. In this contribution, we show how machine-learning techniques can be used to predict Raman spectra based on crystallographic information files (CIF). We are evaluating graph neural networks (GNN), as shown in Figure 1, and foundation model architectures and we are comparing their performance in spectrum prediction tasks. With a dataset containing 600 CIF files/Raman spectrum entries, we achieved a spectrum prediction accuracy above 80%. For scaled applications, we discuss how a foundation model which is pre-trained on large CIF data sets is fine-tuned for Raman spectra prediction. Considering the large amount of CIF files available, scaling AI-assisted spectrum prediction could potentially enable time and cost advantages of Raman spectroscopic analysis in materials discovery.