Inherent structural descriptors via machine learning.

Area of Science:

• Computational chemistry
• Materials science
• Biophysics

Background:

• Identifying collective variables is crucial for interpreting complex system simulations and applying enhanced sampling techniques.
• Current methods face challenges in distilling physically relevant variables from simulation data.

Purpose of the Study:

• To develop a machine learning (ML) approach for identifying physically relevant collective variables from system configurations.
• To apply this ML strategy to characterize complex structural transitions in nanoclusters and peptide systems.

Main Methods:

• A novel machine learning (ML) approach associates instantaneous system configurations with inherent structures from liquids theory.
• The method was applied to a 147-atom gold nanocluster system to analyze structural transitions.
• The ML strategy was also tested on conformational rearrangements of the bradykinin peptide.

Main Results:

• The ML-derived inherent-structure variables effectively characterized structural complexity in gold nanoclusters.
• These variables enabled computation of free-energy landscapes and transition rates.
• The approach successfully described non-equilibrium melting and freezing processes and peptide conformational changes.

Conclusions:

• The proposed ML method offers a powerful tool for discovering essential collective variables in complex systems.
• This approach enhances the interpretation of simulation and experimental data across diverse systems like liquids, glasses, and proteins.
• The ML strategy demonstrates broad applicability and potential for advancing molecular simulations and enhanced sampling techniques.