A fuzzy classification framework to identify equivalent atoms in complex materials and molecules

Area of Science:

• Materials Science
• Computational Chemistry
• Machine Learning

Background:

• The local atomic environment dictates an atom's behavior in various structures like molecules, nanoparticles, and solids.
• Identifying equivalent atoms is crucial for material function analysis, experimental interpretation, and computational efficiency.
• Challenges arise in complex materials lacking ideal symmetries or long-range order.

Purpose of the Study:

• To develop a general machine learning framework for automatically identifying groups of equivalent atoms.
• To provide a robust method for classifying atoms even in non-ideal material structures.

Main Methods:

• Representing local atomic environments using high-dimensional Smooth Overlap of Atomic Positions (SOAP) vectors.
• Employing multidimensional scaling to reduce SOAP vector dimensionality.
• Utilizing mean-shift clustering on the embedded representation for fuzzy classification of atom equivalence.

Main Results:

• Demonstrated the framework's effectiveness on simple aromatic molecules.
• Validated the approach using crystalline Palladium (Pd) surface examples.
• Successfully grouped atoms with nearly equivalent local environments, accounting for thermal vibrations.

Conclusions:

• The proposed machine learning framework offers an automated and generalizable solution for identifying equivalent atoms.
• This method enhances the efficiency and accuracy of atomic-scale modeling and simulation.
• The approach is applicable to diverse material systems, including those with broken symmetries.