Deep learning collective variables from transition path ensemble.

Area of Science:

• Computational Chemistry
• Molecular Dynamics Simulations
• Machine Learning in Science

Background:

• Simulating rare transitions between metastable states is computationally challenging.
• Identifying slow modes, or collective variables, is crucial for efficient molecular dynamics.
• Machine learning methods, like Deep Targeted Discriminant Analysis, have emerged to learn collective variables.

Purpose of the Study:

• To improve the accuracy and convergence of molecular dynamics simulations for rare events.
• To enhance the Deep Targeted Discriminant Analysis (DTDA) method by incorporating transition path data.
• To develop more effective collective variables for studying molecular dynamics.

Main Methods:

• Enriching the training dataset for Deep Targeted Discriminant Analysis (DTDA) with transition path ensemble data.
• Utilizing reactive trajectories generated by On-the-fly Probability Enhanced Sampling (OPEPS) flooding.
• Training new collective variables using an expanded dataset.

Main Results:

• The newly trained collective variables demonstrate improved accuracy in molecular dynamics simulations.
• Faster convergence rates were observed when using the enhanced collective variables.
• The performance was validated on representative molecular systems.

Conclusions:

• Incorporating transition path data significantly enhances machine learning-based collective variables.
• The improved collective variables offer a more efficient approach to simulating rare events in molecular dynamics.
• This methodology advances the capability of computational chemistry for complex molecular systems.