Data sampling scheme for reproducing energies along reaction coordinates in high-dimensional neural network...

Area of Science:

• Computational Chemistry
• Materials Science
• Machine Learning

Background:

• Accurate prediction of reaction pathways is crucial in chemistry.
• High-dimensional neural network potentials offer a computationally efficient alternative to traditional methods like density functional theory.
• Developing efficient data sampling strategies is key to improving the accuracy of these potentials.

Purpose of the Study:

• To propose and validate a new data sampling scheme for high-dimensional neural network potentials.
• To enable accurate prediction of energies along chemical reaction pathways.
• To assess the scheme's performance across diverse chemical reactions.

Main Methods:

• Development of a hybrid data sampling scheme.
• Integration of partial geometry optimization for intermediate structures.
• Inclusion of random atomic displacement for enhanced sampling.
• Calculation of reaction pathway energies using hybrid density functional theory (DFT).

Main Results:

• The proposed data sampling scheme successfully predicted energies along reaction pathways.
• High accuracy was achieved for five distinct chemical reactions, including Claisen rearrangement and Diels-Alder reactions.
• The scheme demonstrated robustness in handling complex reaction mechanisms.

Conclusions:

• The combined partial geometry optimization and random displacement sampling is effective for high-dimensional neural network potentials.
• This approach significantly enhances the predictive accuracy of energies along reaction pathways.
• The validated scheme provides a reliable method for computational studies of chemical reactivity.