Predicting redox potentials by graph-based machine learning methods

Area of Science:

• Computational Chemistry
• Electrochemistry
• Machine Learning

Background:

• Accurate prediction of oxidation and reduction potentials is crucial in chemistry.
• Theoretical computations are often resource-intensive and time-consuming.
• Machine learning offers a promising alternative for efficient potential prediction.

Purpose of the Study:

• To apply machine learning, focusing on graph-based methods, for predicting oxidation and reduction potentials.
• To introduce the ORedOx159 database for evaluating these methods.
• To demonstrate improved accuracy and efficiency in computational electrochemistry.

Main Methods:

• Development of the ORedOx159 database with 318 reactions and 159 organic compounds.
• Review of graph-based machine learning techniques (graph edit distances, kernels, neural networks).
• Assessment of machine learning model performance using fast-computed descriptors.

Main Results:

• Machine learning models achieved notable prediction accuracy for potentials.
• Mean Absolute Error (MAE) of 5.6 kcal/mol for reduction and 7.2 kcal/mol for oxidation potentials.
• Fast descriptor computation significantly improved predictive performance.

Conclusions:

• Machine learning, especially graph-based approaches, provides an efficient route to predict electrochemical potentials.
• The ORedOx159 database serves as a valuable resource for method development.
• This work facilitates the in silico design of novel electrochemical systems.