Unsupervised machine learning in atomistic simulations, between predictions and understanding.

Area of Science:

• Computational chemistry and physics
• Materials science
• Data science

Background:

• Automated analysis of atomistic simulations is essential for linking atomic behavior to collective properties.
• Coarse-graining methods like clustering and dimensionality reduction simplify complex system dynamics.
• Machine learning (ML) has become integral to atomistic modeling, offering diverse data-driven approaches.

Purpose of the Study:

• To review unsupervised machine learning methods for classification and coarse-graining in molecular simulations.
• To examine the mathematical underpinnings of ML techniques in this context.
• To discuss the importance of data representation and potential biases in ML analyses.

Main Methods:

• Review of unsupervised machine learning algorithms (clustering, dimensionality reduction).
• Analysis of data-driven approaches for molecular simulations.
• Comparison of unsupervised and supervised ML techniques for structure-property relations.

Main Results:

• Unsupervised ML methods offer powerful tools for classifying and coarse-graining molecular simulation data.
• Concise atomic structure representations are vital for accurate analysis.
• Supervised ML methods are less prone to bias when predicting material properties from structure.

Conclusions:

• Integrating supervised and unsupervised ML approaches is key to unlocking their full potential in molecular and materials science.
• Understanding the relationship between data analysis frameworks and physical principles is paramount.
• Careful consideration of biases is necessary when interpreting ML-driven structure-property relationships.