Machine-Learning Predictions of Rate Constants of Internal Conversion Using Electronic and Structural Descriptors.

Area of Science:

• Computational chemistry
• Photophysics
• Machine learning

Background:

• Internal conversion (IC) is a key photophysical process, deactivating excited electronic states.
• Accurate prediction of IC rate constants (k_IC) is crucial for understanding molecular behavior but computationally challenging.

Purpose of the Study:

• Develop a machine learning (ML) approach to predict k_IC between the first excited singlet state (S1) and ground state (S0).
• Utilize molecular electronic and structural descriptors for accurate k_IC estimation.
• Provide a cost-efficient alternative to traditional quantum mechanical calculations.

Main Methods:

• Generated a dataset of over 5000 k_IC values using a cost-efficient energy transfer model.
• Employed time-dependent density functional theory (TDDFT) for parameter calculations.
• Trained CatBoost (CB) and neural network (NN) models, including transformer and graph-based architectures.

Main Results:

• Achieved high prediction accuracy with R² values around 0.99 for both CB and NN models.
• Demonstrated successful prediction of k_IC for novel molecules using energy-augmented datasets.
• Introduced novel graph-based X-H bond descriptors for predicting rate constants across diverse chemical classes.

Conclusions:

• The developed ML models offer a highly accurate and efficient method for predicting k_IC.
• Molecular electronic and structural descriptors, particularly graph-based X-H bond descriptors, are effective predictors of photophysical properties.
• This approach significantly reduces the computational cost associated with determining photophysical properties.