Improving Vibrational Mode Interpretation Using Bayesian Regression

To streamline the interpretation of vibrational spectra, this work introduces the use of Bayesian linear regression with automatic relevance determination as a viable approach to decompose the atomic motions along any vibrational mode as a weighted combination of displacements along chemically meaningful internal coordinates. This novel approach denominated vibrational mode automatic relevance determination (VMARD) is presented and compared with the well-established potential energy decomposition (PED) scheme. Good agreement is generally attained between the two methods. VMARD returns a decomposition of the atomic displacement using only a small number of internal coordinates, thus aiding the interpretation of the vibrational spectra. Moreover, the results show that the VMARD descriptions are resilient toward the addition of additional internal coordinates, achieving a concise description of the vibrational modes despite the use of redundant internal coordinates. Potential applications of VMARD involving the gathering of physical insights on the atomic motions along the reaction coordinate at transition state structures, as well as the improvement of theoretically predicted vibrational frequencies, are also presented under a proof-of-concept perspective.