A Multiscale Method for Two-Component, Two-Phase Flow with a Neural Network Surrogate

TL;DR

This paper introduces a multiscale method combining continuum and molecular dynamics models, enhanced with a neural network surrogate, to simulate complex two-phase flow dynamics with microscale phase changes.

Contribution

It presents a novel multiscale approach integrating neural network surrogates for molecular dynamics to accurately simulate two-component, two-phase flow dynamics.

Findings

Successfully simulated droplet dynamics of argon-methane mixtures.

Demonstrated the effectiveness of neural network surrogates in reducing computational complexity.

First to model microscale phase-change transfer rates in compressible two-phase flows.

Abstract

Understanding the dynamics of phase boundaries in fluids requires quantitative knowledge about the microscale processes at the interface. We consider the sharp-interface motion of compressible two-component flow, and propose a heterogeneous multiscale method (HMM) to describe the flow fields accurately. The multiscale approach combines a hyperbolic system of balance laws on the continuum scale with molecular-dynamics simulations on the microscale level. Notably, the multiscale approach is necessary to compute the interface dynamics because there is -- at present -- no closed continuum-scale model. The basic HMM relies on a moving-mesh finite-volume method, and has been introduced recently for compressible one-component flow with phase transitions in [Magiera and Rohde, JCP. 469 (2022)]. To overcome the numerical complexity of the molecular-dynamics microscale model a deep neural network is employed as an efficient surrogate model. The entire approach is finally applied to simulate droplet dynamics for argon-methane mixtures in several space-dimensions. Up to our knowledge such compressible two-phase dynamics accounting for microscale phase-change transfer rates have not yet been computed.