Exploring order parameters and dynamic processes in disordered systems via variational autoencoders.

Area of Science:

• Materials Science
• Data Science
• Chemistry

Background:

• Analyzing dynamic atomic-scale transformations with incomplete data is challenging.
• Understanding structural and chemical changes requires robust data interpretation methods.

Purpose of the Study:

• To develop a novel approach for bottom-up system description from dynamic, atomically resolved imaging data, even with partial or uncertain atomic positions.
• To create a method that effectively reduces complex data while retaining maximum information for analyzing dynamic processes.

Main Methods:

• Utilizing the parsimony of physical descriptors and rotational invariance of noncrystalline solids.
• Implementing a rotationally invariant extension of the variational autoencoder on semantically segmented atom-resolved data.
• Applying the approach to study electron beam-induced processes in silicon-doped graphene.

Main Results:

• Successfully explored the dynamic evolution of electron beam-induced processes in silicon-doped graphene.
• Developed a method to find the most effective reduced representation of atomic systems from imaging data.
• Demonstrated the capability to handle partial or uncertain atomic position data.

Conclusions:

• The proposed approach provides a powerful tool for bottom-up description of systems with large-scale structural and chemical changes.
• This method can be broadly applied to various atomic-scale and mesoscopic phenomena, enabling the introduction and exploration of order parameters.
• It facilitates the study of dynamic evolution in response to external stimuli.