[Paper Review] Enhanced Sampling with Machine Learning: A Review

Mehdi Shams, Zachary A. Smith+3 more|arXiv (Cornell University)|Jun 15, 2023

Machine Learning in Materials Science9 citations

TL;DR

This is a survey of how machine learning is integrated with enhanced sampling in molecular dynamics, covering dimensionality reduction, adaptive strategies, and flow-based free-energy estimation, with discussion of open problems.

ABSTRACT

Molecular dynamics (MD) enables the study of physical systems with excellent spatiotemporal resolution but suffers from severe time-scale limitations. To address this, enhanced sampling methods have been developed to improve exploration of configurational space. However, implementing these is challenging and requires domain expertise. In recent years, integration of machine learning (ML) techniques in different domains has shown promise, prompting their adoption in enhanced sampling as well. Although ML is often employed in various fields primarily due to its data-driven nature, its integration with enhanced sampling is more natural with many common underlying synergies. This review explores the merging of ML and enhanced MD by presenting different shared viewpoints. It offers a comprehensive overview of this rapidly evolving field, which can be difficult to stay updated on. We highlight successful strategies like dimensionality reduction, reinforcement learning, and flow-based methods. Finally, we discuss open problems at the exciting ML-enhanced MD interface.

Motivation & Objective

• Motivate the integration of machine learning with enhanced MD to address time-scale limitations.
• Provide a comprehensive overview of ML-enabled strategies for bias deposition, adaptive sampling, and generalized ensembles.
• Summarize state-of-the-art ML methods used to identify low-dimensional RCs and accelerate sampling.
• Highlight open problems and future directions at the ML-enhanced MD interface.

Proposed method

• Classify enhanced sampling into biasing methods, adaptive sampling, and generalized ensemble methods.
• Discuss dimensionality reduction techniques for identifying slow modes and reaction coordinates (RCs).
• Describe transfer-operator-based approaches (e.g., VAC, TICA, kTICA) and their neural-network extensions (SRV, VAMPnets, GREST).
• Explain information-bottleneck and generative-model-based approaches (MESA, time-lagged autoencoders, RAVE, SPIB) for learning RCs and guiding simulations.
• Explain how reweighting and bias-correction (e.g., Girsanov reweighting) are incorporated to learn from biased trajectories.
• Outline integration with enhanced sampling frameworks and strategies to iteratively improve RCs.

Experimental results

Research questions

• RQ1 How can ML identify or construct low-dimensional RCs that capture slow dynamics in MD?
• RQ2 What ML strategies best support adaptive and generalized-ensemble sampling to overcome energy barriers?
• RQ3 How can biases introduced during sampling be accounted for and corrected in learning RCs and free energies?
• RQ4 What are the trade-offs between model complexity, interpretability, and sampling efficiency in ML-enhanced MD?
• RQ5 What open problems remain at the interface of ML and enhanced MD and how can they be addressed?

Key findings

• ML methods enable automated CV selection and RC learning, reducing manual effort in enhanced sampling.
• Transfer-operator-based and variational approaches (VAC, TICA, SRV) can be implemented with neural networks to learn slow modes.
• Generative and time-lagged models (MESA, time-lagged autoencoders, RAVE, SPIB) enable learning RCs that reflect biased dynamics and improve iterative sampling.
• Girsanov reweighting and related dynamical corrections enable proper averaging over biased path ensembles.
• Hybrid strategies combining multiple enhanced-sampling classes and ML analyses provide more robust exploration of configurational space.
• The review identifies persistent challenges such as model interpretability, RC learning, symmetry exploitation, and reliable free-energy estimation.