Collective variable discovery and enhanced sampling using autoencoders: Innovations in network architecture and error...

Area of Science:

• Computational Chemistry
• Machine Learning in Molecular Dynamics
• Data-Driven Discovery

Background:

• Auto-associative neural networks (autoencoders) are effective for nonlinear dimensionality reduction in molecular simulations.
• They provide explicit, differentiable collective variables suitable for enhanced sampling techniques.
• Accelerated exploration of molecular configurational space is crucial for understanding complex systems.

Purpose of the Study:

• To enhance autoencoder architectures for improved collective variable discovery.
• To generalize the integration of collective variable discovery with enhanced sampling methods.
• To facilitate the accelerated exploration of molecular configurational space.

Main Methods:

• Development of sophisticated neural network architectures, including circular nodes for periodicity and hierarchical structures for variable ranking.
• Implementation of generalized encoder-decoder architectures with bespoke error functions for prior knowledge integration.
• Application of an open-source plugin for the OpenMM molecular simulation package.

Main Results:

• Demonstrated blind collective variable discovery for alanine dipeptide and Trp-cage.
• Successfully integrated collective variable discovery with enhanced sampling techniques.
• Validated the improved performance and generalizability of the refined autoencoder approach.

Conclusions:

• The refined autoencoder architectures significantly improve the process of interleaved collective variable discovery and enhanced sampling.
• This approach offers a powerful tool for analyzing molecular simulation data and exploring complex free energy landscapes.
• The developed methods and open-source plugin enable broader application in computational chemistry and molecular dynamics.