Automated and Efficient Sampling of Chemical Reaction Space.

Area of Science:

• Computational Chemistry
• Materials Science
• Machine Learning

Background:

• Machine learning interatomic potentials (MLIPs) offer high accuracy at computational speeds comparable to classical force fields.
• The performance of MLIPs is critically dependent on the quality and diversity of their training datasets.
• A significant bottleneck in MLIP development is the lack of automated, efficient methods for sampling chemical reaction space.

Purpose of the Study:

• To develop an efficient and fully automated approach for sampling chemical reaction space for MLIP training data generation.
• To address the gap in current MLIP development by creating diverse datasets that capture equilibrium and reactive regions of potential energy surfaces.
• To systematically explore underrepresented reaction pathways and transition states in MLIP training sets.

Main Methods:

• Combining fast tight-binding calculations with selective high-level electronic structure refinement.
• Utilizing single-ended growing string and nudged elastic band methods to explore reaction pathways.
• Systematically generating diverse datasets covering equilibrium and reactive molecular configurations.

Main Results:

• Generation of diverse datasets crucial for robust MLIP development.
• Systematic exploration of reaction pathways, particularly near transition states.
• Creation of datasets with rich structural and chemical diversity.

Conclusions:

• The presented automated workflow enables efficient generation of high-quality training data for MLIPs.
• This approach enhances the accuracy and reliability of MLIPs by covering critical regions of potential energy surfaces.
• The open-source code facilitates the integration of this data generation strategy into existing MLIP development pipelines.