Mapping Simulated Two-Dimensional Spectra to Molecular Models Using Machine Learning.

Area of Science:

• Physical Chemistry
• Computational Chemistry
• Spectroscopy

Background:

• Two-dimensional (2D) spectroscopy provides rich data on molecular properties and dynamics.
• Interpreting complex 2D spectra to extract chemical information is a significant challenge.
• Existing methods may struggle to fully decode the relationship between spectral features and molecular Hamiltonians.

Purpose of the Study:

• To investigate the capability of feedforward neural networks (NNs) in mapping simulated 2D electronic spectra to physical Hamiltonians.
• To assess the accuracy of NNs in characterizing Hamiltonian parameters for molecular systems.
• To demonstrate a computational approach for enhancing the interpretation of 2D spectroscopic data.

Main Methods:

• Generation of hundreds of simulated 2D electronic spectra for monomers and dimers.
• Inclusion of varied Franck-Condon active vibrations and monomer-monomer electronic couplings in simulations.
• Training and evaluation of simple feedforward neural networks (NNs) to predict Hamiltonian parameters from spectral data.

Main Results:

• NNs demonstrated high accuracy, exceeding 90%, in characterizing most Hamiltonian parameters studied.
• The models successfully mapped simulated 2D spectra to their corresponding underlying physical Hamiltonians.
• The study confirmed the potential of NNs in quantitative spectral analysis.

Conclusions:

• Feedforward neural networks are effective tools for interpreting 2D electronic spectroscopy data.
• NNs can bridge the gap between complex spectral features and fundamental molecular parameters (effective Hamiltonians).
• This work highlights a promising computational strategy for advancing chemical dynamics and property analysis from spectroscopic experiments.