Neural networks-based variationally enhanced sampling.

Area of Science:

• Computational Chemistry
• Materials Science
• Statistical Mechanics

Background:

• Atomistic simulations face challenges sampling complex free-energy surfaces due to kinetic bottlenecks.
• Direct simulation approaches are often ineffective for systems with significant energy barriers.
• Accelerating sampling typically involves identifying collective variables and applying bias potentials.

Purpose of the Study:

• To propose a novel method for determining bias potentials using machine learning.
• To integrate machine learning with the variationally enhanced sampling method.
• To accelerate the sampling of complex free-energy landscapes in atomistic simulations.

Main Methods:

• Utilized machine learning, specifically neural networks, to represent the bias potential.
• Employed a variational learning scheme to determine neural network parameters by minimizing a functional.
• Developed an efficient minimization technique to optimize the learning process.

Main Results:

• Neural network expressivity effectively represents complex and rapidly varying free-energy surfaces.
• The proposed method mitigates artifacts caused by boundary effects in simulations.
• The approach successfully handles multiple collective variables for enhanced sampling.

Conclusions:

• Machine learning, integrated with variationally enhanced sampling, provides an effective strategy for determining bias potentials.
• This approach significantly accelerates the exploration of complex free-energy landscapes in atomistic simulations.
• The method offers a powerful tool for overcoming sampling limitations in computational studies.