Multi-Solvent Graph Neural Network for Reduction Potential Prediction Across the Chemical Space

Area of Science:

• Computational chemistry
• Materials science
• Machine learning

Background:

• Redox potentials are crucial for applications like catalysis and energy storage.
• Solvent environments significantly alter molecular redox properties.
• Predicting these potentials accurately across solvents is challenging.

Purpose of the Study:

• To develop a predictive model for redox potentials of organic molecules.
• To enable accurate predictions across diverse solvent environments.
• To facilitate inverse design of molecules with target redox potentials.

Main Methods:

• A message passing graph neural network with Set Transformer readout was employed.
• The model was trained on ~20,000 reduction potentials from the ReSolved dataset.
• Density functional theory (DFT) was used for rigorous computation of potentials.

Main Results:

• The predictor model achieved high accuracy with a mean absolute error of ~0.2 eV.
• The model demonstrated generalization to previously unseen solvents.
• An evolutionary algorithm was coupled for inverse design.

Conclusions:

• The developed graph neural network accurately predicts redox potentials across solvents.
• This approach enables efficient inverse design of redox-active molecules for specific applications.
• The findings have implications for catalyst, antioxidant, and electrode material development.