Unsupervised learning of representative local atomic arrangements in molecular dynamics data

Area of Science:

• Computational chemistry
• Materials science
• Data science

Background:

• Molecular dynamics (MD) simulations generate vast datasets, posing significant data-mining challenges.
• Human interpretation of MD data can be limited or biased, potentially missing critical information.
• Effective analysis methods are crucial for extracting meaningful insights from complex simulation outputs.

Purpose of the Study:

• To develop a quantitative method for characterizing prevalent coordination environments in MD data.
• To overcome the limitations of human interpretation in analyzing large simulation datasets.
• To reveal detailed cation coordination in molecular liquid electrolytes.

Main Methods:

• Combining dimensionality reduction (UMAP) with unsupervised hierarchical clustering (HDBSCAN).
• Extracting distinct molecular formulas within coordination spheres to reduce data complexity.
• Utilizing alignment or shape-matching algorithms to partition formulas into structural isomer families.

Main Results:

• Successfully characterized prevalent coordination environments in MD data.
• Quantitatively identified and classified structural isomer families based on local coordination.
• Revealed detailed insights into cation coordination in molecular liquid electrolytes.

Conclusions:

• The developed method offers an efficient and quantitative approach to analyze MD simulation data.
• This technique enhances the discovery of critical information often missed by traditional analysis.
• The findings contribute to a deeper understanding of chemical species coordination in complex systems like electrolytes.