Efficient sampling of high-energy states by machine learning force fields.

Area of Science:

• Molecular sciences
• Computational chemistry
• Materials science

Background:

• Machine learning (ML) methods offer a powerful approach to bridge the accuracy of ab initio calculations with the efficiency of classical molecular mechanics.
• ML algorithms reconstruct potential energy surfaces from reference data, creating computationally inexpensive force fields for molecular systems.
• The accuracy of ML force fields is critically dependent on the training dataset's quality and coverage.

Purpose of the Study:

• To address the limitations of standard ML training datasets that may omit high-energy states.
• To propose and validate a novel method for generating ML input data that ensures accuracy across the entire configurational phase space.
• To improve the reliability of ML force fields for molecular simulations.

Main Methods:

• Development of a biased subsampling strategy for generating ML training data.
• Targeted inclusion of configurations from high-energy states and energy barriers.
• Application of the method to conformational rearrangements in flexible heterocyclic molecules.

Main Results:

• Demonstration that omitting high-energy states can lead to significant accuracy loss in ML force fields.
• Validation of the biased subsampling approach for enhancing the representation of critical configurational regions.
• Successful application to conformational analysis of heterocyclic systems, improving force field accuracy.

Conclusions:

• Biased subsampling of configurations is essential for creating robust ML force fields that are accurate across the entire configurational phase space.
• This method overcomes the limitations of datasets following Boltzmann distribution, particularly for systems with energy barriers.
• The proposed data generation strategy is a key component for developing reliable and broadly applicable ML force fields in molecular sciences.