An iterative and automatic collective variable optimization scheme via unsupervised feature selection with CUR matrix...

Area of Science:

• Computational Physics and Chemistry
• Materials Science
• Statistical Mechanics

Background:

• Phase transitions often involve high energy barriers, requiring specialized techniques for molecular dynamics simulations.
• Optimizing collective variables (CVs) for enhanced sampling is challenging, especially with limited prior knowledge of the system.
• Accurate characterization of transition pathways is crucial for understanding material properties under extreme conditions.

Purpose of the Study:

• To develop an unsupervised approach for optimizing collective variables (CVs) in molecular dynamics simulations.
• To enable efficient exploration of high-energy structures and phase-transition pathways without prior knowledge.
• To validate the method using a model system of ultra-high-pressure hydrogen.

Main Methods:

• Iterative application of principal component analysis (PCA) on representative feature variables.
• Feature variable generation using the CUR method for efficient feature space contraction.
• Construction of CVs from simulated x-ray diffraction intensity spectra.

Main Results:

• The unsupervised approach demonstrated self-correction capabilities in identifying probable phase-transition pathways.
• Free energy profiles for phase transitions were accurately reproduced using unbiased molecular dynamics simulations.
• The method successfully characterized transition pathways in a hypothetical three-phase model of ultra-high-pressure hydrogen.

Conclusions:

• The developed unsupervised method effectively optimizes CVs for molecular dynamics simulations of phase transitions.
• The approach shows potential for highly autonomous exploration of complex systems with unknown physical mechanisms.
• This work advances the ability to study elusive physical phenomena through computational modeling.