Neural network approach for predicting infrared spectra from 3D molecular structure

Accurately predicting infrared (IR) spectra in computational chemistry using ab initio methods remains a challenge.Current approaches often rely on an empirical approach or on tedious anharmonic calculations, mainly adapted to semi-rigid molecules.This limitation motivates us to explore alternative methodologies.Previous studies explored machine-learning techniques for potential and dipolar surface generation, followed by IR spectra calculation using classical molecular dynamics.However, these methods are computationally expensive and require molecule-by-molecule processing.Our article introduces a new approach to improve IR spectra prediction accuracy within a significantly reduced computing time.We developed a machine learning (ML) model to directly predict IR spectra from three-dimensional (3D) molecular structures.The spectra predicted by our model significantly outperform those from density functional theory (DFT) calculations, even after scaling.In a test set of 200 molecules, our model achieves a Spectral Information Similarity Metric (SIS) of 0.92, surpassing the value achieved by DFT scaled frequencies, which is 0.57.Additionally, our model considers anharmonic effects, offering a fast alternative to laborious anharmonic calculations.Moreover, our model can be used to predict various types of spectra (Ultraviolet or Nuclear Magnetic Resonance for example) as a function of molecular structure.All it needs is a database of 3D structures and their associated spectra.