Supervised learning for accurate mesoscale simulations of suspension flow in wall-bounded geometries

Abstract

Herein, we have employed a supervised learning approach combined with Core-Modified Dissipative Particle Dynamics Simulations (CM-DPD) in order to develop and design a reliable physics-based computational model that will be used in studying confined flow of suspensions. CM-DPD was recently developed and has shown promising performance in capturing rheological behavior of colloidal suspensions; however, the model becomes problematic when the flow of the material is confined between two walls. Wall-penetration by the particles is an unphysical phenomenon that occurs in coarse-grained simulations such as Dissipative Particle Dynamics (DPD) that mostly rely on soft inter-particle interactions. Different solutions to this problem have been proposed in the literature; however, no reports have been given on how to deal with walls using CM-DPD. Due to complexity of interactions and system parameters, designing a realistic simulation model is not a trivial task. Therefore, in this work we have trained a Random Forest (RF) for predicting wall penetration as we vary input parameters such as interaction potentials, flow rate, volume fraction of colloidal particles, and confinement ratio. The RF predictions were compared against simulation tests, and a sufficiently high accuracy and low errors were obtained. This study shows the viability and potentiality of ML combined with DPD to perform parametric studies in complex fluids.