Machine Learning for Molecular Simulation.

Area of Science:

• Scientific computing
• Computational chemistry
• Materials science

Background:

• Molecular simulations involve complex, time-consuming calculations.
• Machine learning (ML) methods are increasingly applied to scientific problems.
• Existing ML approaches have already impacted molecular simulations.

Purpose of the Study:

• To review recent ML methods applied to molecular simulations.
• To focus on deep neural networks for predicting quantum-mechanical energies and forces.
• To discuss ML for coarse-grained dynamics, free energy calculations, and generative sampling.

Main Methods:

• Review of (deep) neural networks for energy and force prediction.
• Application of ML to coarse-grained molecular dynamics.
• Use of ML for free energy surface and kinetics extraction.
• Generative network approaches for sampling molecular structures and thermodynamics.

Main Results:

• ML significantly enhances the efficiency of molecular simulations.
• Deep neural networks provide accurate predictions of quantum-mechanical energies and forces.
• ML facilitates the extraction of thermodynamic and kinetic properties.
• Generative models enable effective sampling of molecular equilibrium structures.

Conclusions:

• ML offers powerful tools to overcome limitations in molecular simulations.
• Integrating molecular physics principles into ML is crucial for methodological advancement.
• Open challenges remain at the interface of ML and molecular simulation, requiring further research.