Nina Andrejevic1,Michael Davis2,Maria Chan2,Mingda Li1
Massachusetts Institute of Technology1,Argonne National Laboratory2
Nina Andrejevic1,Michael Davis2,Maria Chan2,Mingda Li1
Massachusetts Institute of Technology1,Argonne National Laboratory2
Raman spectroscopy captures materials’ vibrational properties in the form of highly resolved fingerprints with characteristic peaks. However, connecting Raman spectra to underlying structural and chemical attributes can be nontrivial, and first principles calculations typically incur a high computational cost. In this work, we apply machine learning methods to address the challenges associated with obtaining Raman spectra from comparatively low-dimensional inputs, which consist only of accessible structural information and single-atom properties. Using an objective function inspired by optimal transport, we first train an autoencoder to obtain learned, low-dimensional representations of Raman spectra. These representations serve as effective prediction targets for more economical machine learning models that then recover Raman spectra from materials’ structural and chemical attributes. To address the limited data available for training these models, we employ the symmetry-constrained Euclidean neural networks<sup>1</sup>, which have been used to successfully predict phonon densities of states from limited training samples owing to their restricted functional space<sup>2</sup>. We apply our proposed framework to <i>ab initio</i> Raman spectra of 2D materials and further investigate its application to experimental Raman spectra from spectral databases of mineral compounds with varying complexity. Our approach enables rapid prediction of Raman spectra from structures which accelerates the interpretation of Raman spectroscopy data.