A fragment based approach towards curating, comparing and developing machine learning models applied in...

Area of Science:

• Computational chemistry
• Machine learning
• Photochemistry

Background:

• Graph neural networks (GNNs) are increasingly used for predicting molecular properties.
• Photochemistry and photocatalysis are gaining importance as sustainable alternatives.
• Existing GNN models struggle with exciton localization in molecules, causing prediction errors.

Purpose of the Study:

• To develop a molecular fragmentation strategy to improve GNN predictions for photophysical properties.
• To address the limitation of exciton localization in current GNN models.
• To enable comparison of structural diversity in molecular libraries.

Main Methods:

• A novel molecular fragmentation strategy was developed.
• A new database, ALFAST-DB, containing 46,432 adiabatic S₀-T₁ energy gaps was generated.
• A fragment-based delta learning approach was employed and compared with traditional message passing GNNs (MPGNNs).

Main Results:

• The fragmentation strategy effectively overcomes exciton localization limitations.
• The fragment-based delta learning approach demonstrated improved model generalizability.
• The new method achieved prediction accuracies comparable to MPGNN architectures.

Conclusions:

• Molecular fragmentation is a viable strategy to enhance GNN performance in photophysical property prediction.
• This approach offers a way to assess and compare the diversity of molecular datasets.
• The developed method holds promise for advancing sustainable chemistry through improved computational modeling.