Exploiting Machine Learning to Efficiently Predict Multidimensional Optical Spectra in Complex Environments

Abstract

The excited-state dynamics of chromophores in complex environments determine a range of vital biological and energy capture processes. Time-resolved, multidimensional optical spectroscopies provide a key tool to investigate these processes. Although theory has the potential to decode these spectra in terms of the electronic and atomistic dynamics, the need for large numbers of excited-state electronic structure calculations severely limits first-principles predictions of multidimensional optical spectra for chromophores in the condensed phase. Here, we leverage the locality of chromophore excitations to develop machine learning models to predict the excited-state energy gap of chromophores in complex environments for efficiently constructing linear and multidimensional optical spectra. By analyzing the performance of these models, which span a hierarchy of physical approximations, across a range of chromophore− environment interaction strengths, we provide strategies for the construction of machine learning models that greatly accelerate the calculation of multidimensional optical spectra from first principles.

Machine Learning for Electronically Excited States of Molecules

“…This model could reproduce reference linear absorption spectra, spectral densities, and could capture spectral diffusion of two-dimensional electronic spectra of all treated chromophores. 539 …”

“…About two-thirds rely on NNs, whereby simple multilayer feed-forward NNs are most often employed. Several research fields were advanced with NN-fitted functions: photodynamics simulations (refs ( 15 , 93 − 95 , 98 , 140 , 141 , 143 , 144 , 147 , 392 , 412 , and 413 )), spectra predictions and analysis, 97 , 152 , 155 , 241 , 490 , 510 , 539 excited-state properties, 15 − 17 , 92 , 95 , 97 , 510 diabatization procedures, 96 , 142 interpretation of reaction outcomes, 160 , 696 and the prediction of HOMO–LUMO gaps or gaps between energetic states. 77 , 152 , 158 In these studies, between one and seven hidden layers with varying numbers of nodes were used.…”

Molecular excited states through a machine learning lens

“…Another challenge for ML is the proper description of environmental effects (such as solvents) on excitation energies and two possible strategies have been suggested to address this problem. 34 One approach would be to completely ignore the environment in the ML model and only incorporate environmental effects implicitly through training on reference data that include such effects. This approach is equivalent to implicit solvent models in QM.…”

“…34 Alternatively, information about the environment can be included explicitly in the ML model, typically at the descriptor level. 34,122 In a broader context, it is known from the use of ML approaches for the study of ground-state properties that special treatment of long-range interactions, such as explicit inclusion of additional dispersion corrections or training separate ML models on charges for capturing electrostatic effects, is often necessary. 23,123 In any case, the generation of reference data that capture environmental effects requires additional computational cost.…”

“…ML is emerging as a useful tool for spectroscopy and can assist in both predicting and analyzing the electronic spectrum for a given chemical structure and solving the inverse problem of determining a chemical structure for a given spectrum. A wide range of electronic spectroscopic methods exist, which measure various physicochemical phenomena, 9 and ML has been already applied to assist different spectroscopic techniques, including steady-state absorption (in both optical 26,32,34,35,125,126 and X-ray 127-129 domains), emission, 28 and multidimensional time-resolved spectroscopy. 34,126,130 Some of these studies [26][27][28] were the earliest works using ML for excited-state research, and we are currently experiencing a resurgence in interest in such methods.…”