An extended autoencoder model for reaction coordinate discovery in rare event molecular dynamics datasets

Area of Science:

• Computational Chemistry
• Molecular Dynamics
• Statistical Mechanics

Background:

• The reaction coordinate (RC) is vital for understanding molecular processes in simulations.
• Current methods for identifying RCs require manual selection of collective variables (CVs).
• The committor function is the optimal RC but is high-dimensional and difficult to interpret.

Purpose of the Study:

• To automate the dimensionality reduction of RCs using an extended autoencoder.
• To optimize a low-dimensional latent space for both data reconstruction and committor prediction.
• To provide a more interpretable and accurate representation of molecular processes.

Main Methods:

• Developed an extended autoencoder model for dimensionality reduction.
• Mapped high-dimensional CVs to a lower-dimensional latent space.
• Used the latent space for reconstruction and committor function prediction.

Main Results:

• The extended autoencoder effectively extracts reaction mechanisms.
• Achieved reliable predictions of the committor function.
• Demonstrated the model's utility on toy systems and methane hydrate nucleation data.

Conclusions:

• The extended autoencoder provides an automated and effective approach to identify optimal low-dimensional RCs.
• This method offers valuable atomistic insights into complex molecular processes.
• The model shows potential for generating reaction paths and improving molecular simulations.