Size-Transferable Prediction of Excited State Properties for Molecular Assemblies with a Machine Learning Exciton...

Area of Science:

• Computational chemistry
• Quantum mechanics
• Machine learning applications

Background:

• Computational modeling of excited states in molecular aggregates is computationally intensive.
• Existing machine learning models lack scalability due to training on specific aggregate sizes.

Purpose of the Study:

• To develop a scalable machine learning approach for modeling excited states of molecular aggregates.
• To address computational challenges and size heterogeneity in aggregate modeling.

Main Methods:

• Decomposing the exciton model Hamiltonian of large aggregates into dimer pairs.
• Training a machine learning model on these dimer pairs to reconstruct Hamiltonians for any aggregate size.
• Implementing a new phase-correction method using approximations for coupling terms.

Main Results:

• Accurate prediction of excitation energies for perylene and tetracene trimers and tetramers.
• Estimation of S1 oscillator strengths for perylene aggregates.
• Analysis of optical gaps in nanosized perylene aggregates (up to 50 monomers) revealing coupling effects on size dependency.

Conclusions:

• A dimer-based machine learning model enables scalable Hamiltonian reconstruction for molecular aggregates of any size.
• The developed method accurately predicts excited-state properties and provides insights into aggregate behavior.
• Future research will focus on transferability across different monomers for heterogeneous assemblies.