Data-Efficient Active Learning for Thermodynamic Integration: Acidity Constants of BiVO₄ in Water

Area of Science:

• Computational Chemistry
• Materials Science
• Physical Chemistry

Background:

• Protonation states are critical for catalysis, geochemistry, biochemistry, and pharmaceutics.
• Accurate determination of acidity constants is challenging due to electronic structure, thermal fluctuations, vibrations, and solvation.
• Existing methods like ab-initio molecular dynamics are computationally expensive for achieving necessary simulation timescales.

Purpose of the Study:

• To develop an efficient and accurate method for calculating acidity constants.
• To investigate deprotonation reactions at the BiVO₄ (010)-water interface for photocatalytic water splitting.
• To overcome the computational limitations of traditional simulation methods.

Main Methods:

• Employed thermodynamic integration accelerated by committee Neural Network potentials.
• Trained a single machine learning model to describe protonated, deprotonated, and intermediate states.
• Investigated deprotonation reactions at the BiVO₄ (010)-water interface.

Main Results:

• Demonstrated convergence of ensemble averages and acidity constants with simulation time and Kirkwood coupling parameter.
• Showed that nanosecond simulation times are necessary for statistical convergence, achievable with the ML approach.
• Highlighted that the ML workflow requires significantly fewer calculations than ab-initio molecular dynamics.

Conclusions:

• The developed machine learning workflow significantly advances the calculation of free energy differences with ab-initio accuracy.
• This approach enables the study of systems requiring extended timescales, previously inaccessible with standard methods.
• The method provides a computationally efficient alternative for determining acidity constants and studying interfacial reactions.