Evidence That Less Can Be More for Transferable Force Fields.

Area of Science:

• Computational Chemistry
• Materials Science
• Molecular Dynamics

Background:

• Graph-based parameter assignment is foundational for transferable force fields in molecular dynamics.
• Force field specificity varies, impacting accuracy, complexity, and training data needs.

Purpose of the Study:

• To quantify the trade-offs between force field specificity and accuracy.
• To evaluate the impact of graph specificity on molecular dynamics simulation performance.

Main Methods:

• Parametrized three sets of force fields with varying graph specificity.
• Utilized a shared procedure for generating quantum-chemically derived training data.
• Benchmarked force fields on structural features and liquid properties of 87 organic molecules across 146 state points.

Main Results:

• Accuracy for trained properties saturates rapidly with increasing graph specificity.
• More complex, less transferable force fields show marginal benefits with standard training data.
• Complex force fields performed slightly worse on properties not included in training.

Conclusions:

• Increasing force field complexity offers limited advantages for accuracy with common training data sources.
• The benefits of highly specific force fields should be carefully weighed against performance gains and data requirements.
• Fortuitous regularization in less specific force fields may explain observed performance trends.