When do short-range atomistic machine-learning models fall short?

Area of Science:

• Computational chemistry
• Materials science
• Machine learning

Background:

• Atomistic machine-learning models are increasingly used in molecular simulations.
• These models excel at learning complex local interactions but often lack explicit long-range interaction descriptions.
• Understanding the necessity and impact of long-range interactions is crucial for model development.

Purpose of the Study:

• To investigate the role of long-range interactions in atomistic machine-learning models.
• To benchmark model performance across different property types (fitting accuracy, cluster properties, bulk thermodynamics).
• To identify the limitations of short-range machine-learning models.

Main Methods:

• Utilized a flexible Extended Simple Point Charge (SPC/E) water model as a benchmark system.
• Analyzed fitting accuracy, isolated cluster properties, and bulk thermodynamic properties.
• Compared performance of models with and without explicit long-range interaction descriptions.

Main Results:

• Local machine-learning representations are adequate for predicting condensed liquid phase properties.
• Short-range models demonstrate limitations in accurately representing cluster and vapor phase properties.
• The inclusion of long-range interactions significantly improves the prediction of these properties.

Conclusions:

• Explicit long-range interactions are essential for accurately modeling cluster and vapor phase properties in atomistic machine-learning models.
• Local machine-learning models have specific regimes where their accuracy is limited.
• This study provides insights into the necessary components for robust machine-learning models in molecular simulations.