Linear Atomic Cluster Expansion Force Fields for Organic Molecules: Beyond RMSE.

Area of Science:

• Computational chemistry
• Materials science
• Molecular dynamics

Background:

• Traditional molecular mechanics force fields offer speed but limited accuracy.
• Machine learning (ML) force fields achieve high accuracy but can be computationally expensive.
• Accurate and efficient force fields are crucial for molecular simulations.

Purpose of the Study:

• To develop fast and accurate molecular force fields using the atomic cluster expansion (ACE) framework.
• To demonstrate the performance of ACE force fields compared to empirical and ML approaches.
• To assess the capabilities of ACE beyond root-mean-square error (RMSE), including transferability and extrapolation.

Main Methods:

• Parametrization of potential energy surfaces using body-ordered symmetric polynomials within the ACE framework.
• Validation on benchmark datasets like MD17 and ISO17.
• Comparison with empirical and ML force fields on diverse tasks: normal-mode prediction, high-temperature molecular dynamics, dihedral profile prediction, and bond breaking.

Main Results:

• Four- and five-body ACE force fields achieve accuracy up to 10 times better than empirical force fields.
• ACE models demonstrate accuracy comparable to state-of-the-art ML approaches.
• ACE force fields show excellent performance in predicting molecular properties and simulating complex dynamics, including bond breaking.

Conclusions:

• The ACE framework enables the construction of highly accurate and computationally efficient molecular force fields.
• ACE force fields offer a promising alternative to existing methods, balancing speed and accuracy.
• ACE demonstrates robust transferability and extrapolation capabilities on complex molecular systems.