Learning Efficient, Collective Monte Carlo Moves with Variational Autoencoders.

Area of Science:

• Computational Chemistry
• Molecular Dynamics
• Machine Learning

Background:

• Discovering collective variables is crucial for enhancing molecular simulation sampling.
• Variational autoencoders (VAEs) have been explored for identifying collective variables, but their generative capabilities remain underutilized.

Purpose of the Study:

• To demonstrate the use of VAEs for learning efficient collective Monte Carlo moves.
• To explore the impact of VAE decoding on accelerating molecular simulation sampling.

Main Methods:

• On-the-fly learning of collective variables using VAEs.
• Development of VAE-based Monte Carlo moves for enhanced sampling.
• Analysis of encoding and decoding distribution impact on VAE performance.

Main Results:

• VAE-based Monte Carlo moves significantly accelerate sampling along learned collective variables.
• The proposed method achieves exact sampling without reweighting.
• Acceptance rates of VAE-based moves approach unity in ideal scenarios.

Conclusions:

• VAEs can be effectively utilized to generate efficient collective Monte Carlo moves for molecular simulations.
• The performance of VAE-based moves is sensitive to the fidelity of the decoder in reflecting physical principles.
• This approach offers a powerful, minimally supervised method for enhancing simulation sampling.