Bridging the Experiment-Calculation Divide: Machine Learning Corrections to Redox Potential Calculations in Implicit...

Area of Science:

• Computational chemistry
• Materials science
• Electrochemistry

Background:

• Accurate prediction of redox potentials is crucial for designing materials in catalysis and energy storage.
• Density functional theory (DFT) calculations offer a pathway for rapid predictions but suffer from persistent errors compared to experimental data.
• Addressing these errors is key to advancing electrochemical applications.

Purpose of the Study:

• To develop machine learning (ML) models for reducing errors in DFT-based redox potential calculations.
• To improve predictions in both implicit and explicit solvent models.
• To investigate error sources in explicit solvent redox potential calculations.

Main Methods:

• Developed ML correction models trained on the ROP313 dataset with experimentally measured redox potentials.
• Applied ML to implicit solvent models to reduce systematic bias and outliers.
• Implemented a hybrid implicit-explicit solvent model with GPU-accelerated quantum chemistry for efficient explicit solvent calculations.

Main Results:

• ML correction models significantly reduced errors in redox potential predictions for organic and organometallic compounds.
• ML-corrected potentials showed reduced sensitivity to the choice of DFT functional.
• The combined implicit-explicit solvent model achieved converged results with reduced computational cost, enabling large-scale data generation.

Conclusions:

• Machine learning offers a powerful approach to enhance the accuracy of computational redox potential predictions.
• The developed hybrid solvent model and ML corrections provide a more reliable and efficient method for computational electrochemistry.
• This work paves the way for more accurate computational screening of materials for energy storage and catalysis.