Adaptive enhanced sampling by force-biasing using neural networks.

Area of Science:

• Computational Chemistry
• Machine Learning in Science
• Molecular Dynamics

Background:

• Molecular simulations often face challenges with rugged free energy landscapes.
• Traditional adaptive biasing force methods can struggle with discrete force estimates and sparsely sampled regions.

Purpose of the Study:

• To present a general machine learning-assisted method for molecular simulations.
• To improve sampling efficiency and accuracy in systems with complex free energy landscapes.

Main Methods:

• Utilizing a self-regularizing artificial neural network (ANN) to generate continuous, estimated generalized forces.
• Integrating the ANN within an adaptive biasing force (ABF) framework.
• Applying the method to molecular simulation systems with challenging energy landscapes.

Main Results:

• The ANN-based approach provides smooth force estimates even in sparsely sampled regions.
• The method enables force estimation in previously unexplored areas of the simulation space.
• Continuous force estimates lead to more effective biasing compared to discrete grid-based methods.
• Demonstrated significant enhancements over traditional ABF in three diverse examples.

Conclusions:

• The proposed machine learning-assisted method offers a robust and generalizable approach for molecular simulations.
• It effectively addresses limitations of existing ABF techniques, particularly for complex systems.
• The method shows promise for improving the efficiency and accuracy of molecular simulation studies.