Unsupervised learning for local structure detection in colloidal systems

Area of Science:

• Colloidal science
• Computational physics
• Materials science

Background:

• Identifying local particle environments is crucial for understanding colloidal system behavior.
• Traditional methods often require manual tuning and system-specific parameters.

Purpose of the Study:

• To introduce a simple, fast, and unsupervised algorithm for detecting local environments in colloidal systems.
• To autonomously group similar local environments without prior system knowledge.

Main Methods:

• Utilized standard bond-orientational order parameters to characterize local particle environments.
• Employed a neural network-based autoencoder combined with Gaussian mixture models for environment classification.
• Tested the algorithm on diverse simulated colloidal systems, including mixtures and anisotropic particles.

Main Results:

• Successfully identified relevant local environments across various colloidal systems with high precision.
• Demonstrated effectiveness in analyzing self-assembled structures like fluid-crystal interfaces and grain boundaries.
• The autoencoder identified key bond orientational order parameters relevant to system analysis.

Conclusions:

• The developed unsupervised learning algorithm offers a robust and efficient alternative for local environment detection in colloidal systems.
• This method simplifies analysis and reduces the need for system-specific parameterization.
• Provides insights into the most influential order parameters for characterizing colloidal structures.