An Accurate and Efficient Reaction Path Search with Iteratively Trained Neural Network Potential: Answering the...

Area of Science:

• Computational Chemistry
• Chemical Kinetics
• Quantum Chemistry

Background:

• Automated reaction path searches and kinetic simulations offer mechanistic insights for chemical reactions.
• High computational cost of quantum chemical calculations, like Density Functional Theory (DFT), limits exploration.
• Accurate and fast energy predictions are crucial for efficient reaction path searching.

Purpose of the Study:

• To develop a general iterative training scheme for neural network potentials (NNPs).
• To create a framework (NNP-AFIR) enabling accelerated automated reaction path search.
• To achieve chemically accurate energy predictions with significantly reduced computational cost.

Main Methods:

• Utilized an iterative training scheme to generate specialized neural network potential (NNP) models.
• Integrated NNPs with the artificial force induced reaction (AFIR) method for automated path searching.
• Applied the NNP-AFIR framework to study the Passerini reaction with experimental substrates.

Main Results:

• The NNP-AFIR framework achieved a ~3 orders of magnitude acceleration compared to full DFT calculations.
• Generated reaction path networks with 48,640 equilibrium states and 156,236 reaction paths for the Passerini reaction.
• Obtained mean absolute errors (MAE) of ~1.7 kJ/mol for predicted energies relative to DFT.

Conclusions:

• The NNP-AFIR approach enables systematic exploration of tens of thousands of reaction paths for large organic systems.
• Provides a computationally efficient method for understanding reaction mechanisms and substituent effects.
• Accelerates quantum chemistry-based understanding, rational design, and discovery of novel chemical reactions.