Machine Learning for Video Classification Enables Quantifying Intermolecular Couplings from Simulated Time-Evolved Multidimensional Spectra

Signals in two-dimensional electronic spectroscopy (2DES) encode information about electronic, vibrational, and vibronic couplings in molecular structures. However, chemical information is often difficult to extract. Here, we use a (2+1)-dimensional convolutional neural network ((2+1)D-CNN) to map simulated 2DES spectra to their underlying electronic couplings. The (2+1)D-CNN approach, in contrast to lower-dimensional network architectures, can access all of the time and frequency dimensions in the 2DES signal. We find that the (2+1)D-CNN algorithm classifies regimes of Coulombic couplings in dimers with a 10-fold cross-validation accuracy of (96.2 ± 1.0)%. By examining the optimized filters within the CNN, we find that the (2+1)D-CNN learns from frequency-domain peaks in 2DES spectra and their time evolution (including quantum beating). We also generate and analyze class-activation maps (CAMs) to reveal which features of the spectroscopic data are most important for the (2+1)D-CNN classifications. These studies provide an ML approach to address inverse problems in multidimensional spectroscopy and provide strategies to better understand how chemical information is encoded in spectroscopic data.