Nanoscale slip length prediction with machine learning tools

Filippos Sofos,Theodoros E Karakasidis

doi:10.1038/s41598-021-91885-x

Abstract

This work incorporates machine learning (ML) techniques, such as multivariate regression, the multi-layer perceptron, and random forest to predict the slip length at the nanoscale. Data points are collected both from our simulation data and data from the literature, and comprise Molecular Dynamics simulations of simple monoatomic, polar, and molecular liquids. Training and test points cover a wide range of input parameters which have been found to affect the slip length value, concerning dynamical and geometrical characteristics of the model, along with simulation parameters that constitute the simulation conditions. The aim of this work is to suggest an accurate and efficient procedure capable of reproducing physical properties, such as the slip length, acting parallel to simulation methods. Non-linear models, based on neural networks and decision trees, have been found to achieve better performance compared to linear regression methods. After the model is trained on representative simulation data, it is capable of accurately predicting the slip length values in regions between or in close proximity to the input data range, at the nanoscale. Results also reveal that, as channel dimensions increase, the slip length turns into a size-independent material property, affected mainly by wall roughness and wettability.

Highlights

This work incorporates machine learning (ML) techniques, such as multivariate regression, the multi-layer perceptron, and random forest to predict the slip length at the nanoscale
Nanobubbles have been observed near hydrophobic surfaces[15] and confirmed experimentally[16], in a way that they form a layer that acts as a lubricant, significantly increasing the slip length
Albeit far from replacing simulation methods that have matured over the years in classical physics, chemistry and engineering problems, we show that ML techniques are capable of reproducing fast, accurate prediction of computationally intensive, and, sometimes, ambiguous properties, such as the slip length at the fluid/solid interface, where large temporal fluctuations have been o bserved[32]

Summary

Introduction

This work incorporates machine learning (ML) techniques, such as multivariate regression, the multi-layer perceptron, and random forest to predict the slip length at the nanoscale. Training and test points cover a wide range of input parameters which have been found to affect the slip length value, concerning dynamical and geometrical characteristics of the model, along with simulation parameters that constitute the simulation conditions. In contrast to the macroscale, at this scale confinement effects arise, with significant fluid ordering near the solid surface, non-constant viscosity values, and slip lengths that violate the continuum no-slip assumption[6,7]. It is well-established that there exists a number of interfacial characteristics that affect fluid transport and the degree of slip.

Objectives

Methods

Results

Conclusion