Descriptor-free unsupervised learning method for local structure identification in particle packings.

Area of Science:

• Materials Science
• Computational Physics
• Data Science

Background:

• Local structure identification is crucial in various scientific fields.
• Existing mathematical and supervised learning methods require specific descriptors and are limited to particular packing configurations.
• There is a need for a versatile and descriptor-free method for local structure analysis.

Purpose of the Study:

• To develop an improved unsupervised learning method for descriptor-free local structure identification in particle packing.
• To utilize point cloud data directly from particle spatial positions.
• To enable the discovery of both known and novel local structures in various packing types.

Main Methods:

• An autoencoder architecture was constructed using a point cloud network.
• Gaussian mixture models were integrated for dimension reduction and clustering.
• The method directly processes point cloud data, eliminating the need for predefined descriptors.

Main Results:

• The improved method demonstrated high performance in identifying local structures in quasicrystal disk and sphere packings, comparable to existing techniques.
• For disordered packings, the method identified a previously unknown seven-neighbor motif in maximally dense random disk packing.
• Acentric structural motifs were discovered in random close packing of spheres, showcasing the ability to find unknown structures.

Conclusions:

• The descriptor-free unsupervised learning method offers a powerful tool for analyzing particle packing structures.
• This approach can extract valuable information from large simulation and experimental datasets.
• The method facilitates the development of new order parameters for characterizing particle packings.