Learning Mappings between Equilibrium States of Liquid Systems Using Normalizing Flows

TL;DR

This paper demonstrates how normalizing flows can learn mappings between different liquid systems, enabling unbiased sampling and potential cost-effective simulations at high accuracy.

Contribution

It introduces a novel application of normalizing flows for transforming equilibrium states between liquid systems, facilitating efficient reweighting and simulation.

Findings

Successful transformation between Lennard-Jones systems with different parameters

Effective mapping between Lennard-Jones and repulsive particle systems

Potential for high-accuracy liquid simulations at reduced computational cost

Abstract

Generative models are a promising tool to address the sampling problem in multi-body and condensed-matter systems in the framework of statistical mechanics. In this work, we show that normalizing flows can be used to learn a transformation to map different liquid systems into each other allowing at the same time to obtain an unbiased equilibrium distribution through a reweighting process. Two proof-of-principles calculations are presented for the transformation between Lennard-Jones systems of particles with different depths of the potential well and for the transformation between a Lennard-Jones and a system of repulsive particles. In both numerical experiments, systems are in the liquid state. In future applications, this approach could lead to efficient methods to simulate liquid systems at ab-initio accuracy with the computational cost of less accurate models, such as force field or coarse-grained simulations.