Learning intermolecular forces at liquid-vapor interfaces.

Area of Science:

• Computational chemistry
• Materials science
• Artificial intelligence in physics

Background:

• Artificial neural network (ANN) potentials are increasingly used to model complex materials.
• Describing inhomogeneous and disordered systems, such as liquid-vapor interfaces, presents challenges for ANNs.
• Local representations in ANNs often struggle with long-ranged interactions crucial for interfacial properties.

Purpose of the Study:

• To investigate methods for training ANN potentials to accurately describe inhomogeneous, disordered systems.
• To identify limitations of local ANN potentials and propose strategies for improvement.
• To enhance the description of liquid-vapor interfaces using advanced ANN potential training.

Main Methods:

• Utilized liquid-state theory to inform the training of ANN potentials.
• Compared local ANN potentials with those incorporating explicit models for long-ranged interactions.
• Trained ANNs on short-ranged components while explicitly modeling long-ranged interactions.
• Investigated the role of explicit electrostatics and local molecular field potentials.

Main Results:

• Local ANN potentials accurately describe bulk properties but fail for interfacial properties dependent on unbalanced long-ranged interactions.
• Incorporating explicit models for long-ranged interactions significantly improves the description of interfacial properties.
• Models with explicit electrostatics demonstrate higher accuracy and are easier to train.
• Local ANN models sometimes approximate molecular field potentials, but this is inconsistent.

Conclusions:

• Accurate modeling of inhomogeneous systems requires explicit consideration of long-ranged interactions beyond local atomic environments.
• ANN potentials trained with explicit long-range interaction models robustly capture interfacial phenomena.
• Explicit electrostatic modeling is a promising approach for enhancing ANN potential accuracy in complex systems.