Machine learning short-ranged many-body interactions in colloidal systems using descriptors based on Voronoi cells.

Area of Science:

• Computational physics
• Materials science
• Statistical mechanics

Background:

• Machine learning (ML) accelerates computer simulations for complex systems.
• Capturing many-body interactions in colloidal systems is computationally challenging.
• Realistic simulations require accurate modeling of these intricate interactions.

Purpose of the Study:

• Introduce a novel ML-based strategy for fitting many-body interactions in colloidal systems.
• Develop and apply Voronoi-based descriptors to capture local environments.
• Assess the effectiveness of ML potentials in simulating colloid-polymer mixtures.

Main Methods:

• Developed Voronoi-based descriptors to represent the local environment in colloidal systems.
• Utilized a simple neural network to fit the effective potential.
• Simulated a 2D colloid-polymer mixture with hard-disk like interactions.

Main Results:

• Demonstrated that Voronoi-based descriptors accurately capture the many-body nature of the studied system.
• Found that ML potentials can effectively model complex colloidal interactions.
• Highlighted the insufficiency of Pearson correlation alone for evaluating ML potential predictive power.

Conclusions:

• Voronoi descriptors provide a sufficient and accurate method for capturing many-body interactions in local colloidal systems.
• ML strategies, particularly with appropriate descriptors, significantly advance the simulation of realistic colloidal systems.
• Emphasized the need for comprehensive metrics beyond correlation functions to validate ML-based potentials.