Machine-learning potentials for efficient simulations of anisotropic colloids.

Area of Science:

• Computational physics
• Colloid science
• Materials science

Background:

• Simulating non-spherical colloidal particles presents significant computational challenges due to complex geometric dependencies.
• Existing methods struggle with the intricate force and energy calculations required for arbitrary shapes.

Purpose of the Study:

• To develop and validate a computationally efficient method for simulating interactions between non-spherical colloidal particles.
• To assess the performance of various machine learning models for predicting interaction energies and forces.

Main Methods:

• Representing particle shapes using point clouds instead of position and orientation.
• Comparing descriptor-based (Behler-Parrinello, SO(3)-Equivariant), and end-to-end (SchNet, DimeNet, DimeNet++) models.
• Utilizing Neuroevolution Potential (NEP) for modeling interactions between rigid anisotropic bodies.

Main Results:

• Neuroevolution Potential (NEP) demonstrated an optimal balance of accuracy and computational efficiency.
• NEP accurately reproduced structural properties for diverse shapes (cubes, tetrahedra, bipyramids, twisted cylinders).
• Achieved an order-of-magnitude speedup compared to other simulation methods, with straightforward extension to multi-face shapes.

Conclusions:

• The point cloud representation coupled with NEP enables scalable and accurate simulations of complex colloidal systems.
• This approach facilitates efficient studies on shape-dependent interactions and phase behavior.
• NEP's flexibility and accuracy were confirmed even for asymmetric shapes like twisted cylinders.