Accurate nuclear quantum statistics on machine-learned classical effective potentials.

Area of Science:

• Computational chemistry
• Molecular dynamics
• Machine learning in science

Background:

• Nuclear quantum effects (NQEs) significantly influence hydrogen-bound systems, including biomolecules.
• Simulating NQEs using path integral molecular dynamics (PIMD) is computationally expensive, especially for complex systems at low temperatures.

Purpose of the Study:

• To develop a computationally efficient method for incorporating NQEs in molecular simulations.
• To reduce the computational burden of PIMD by employing machine learning techniques.

Main Methods:

• Leveraging deep learning and multiscale coarse-graining.
• Developing a machine-learned potential to represent corrections to classical potentials.
• Validating the approach on four distinct systems: Morse potential, Zundel cation, single water molecule, and bulk water.

Main Results:

• The machine-learned potential accurately represents corrections to classical potentials, significantly reducing computational cost.
• The framework accurately computes position-dependent static properties, showing excellent agreement with PIMD calculations.
• The approach effectively captures strong NQEs in the tested systems.

Conclusions:

• The developed framework offers a computationally feasible way to simulate NQEs in molecular systems.
• This work paves the way for transferable machine-learned potentials for accurate NQE simulations across diverse molecular systems.