Application of machine learning in colloids transport in porous media studies: Lattice Boltzmann simulation results as training data

Abstract

Colloid transport through a porous medium changes geometrical and hydraulic properties of the pore space. The impact of this effect depends on the colloid types and pore space surface properties which determine the likelihood of pore clogging. Colloid attachment and subsequent detachment are key factors in pore clogging. In this study, the impact of four major fluid and colloids properties on the pore surface and hydraulic conductivity alteration during colloids transport were evaluated using machine learning. These four parameters include solution ionic strength, zeta potential, colloid size and fluid flow velocity. A combined lattice Boltzmann-smoothed profile method was used to simulate accurately coupled mechanisms governing colloid transport to evaluate the impact of the four parameters on the resulting pore space properties after colloid transport. The result of several simulations revealed significant changes of pore surface coverage by the attached colloids, and conductivity, void fraction and coordination number of colloid agglomerates created during transport of individual colloids. Since the simulation of the impact of combination of all possible sets of four parameters is very time consuming, an Artificial Neural Network (ANN) was used as a prognostic method to use the results of several simulations to predict the behavior for a wide range of pore, colloidal and fluid properties. Reported results from a set of 162 simulation case studies for different possible combination of solution ionic strength, zeta potentials, colloid size and flow velocity were selected as input parameters for the machine learning. Four output parameters, namely, pore surface coverage, conductivity, void fraction and coordination number of the colloidal particles were selected. To lower the prediction error value, which is targeted to be lower than 10%, networks were trained 50 times using a MATLAB code, and in each training, after at most 10 epochs, networks were trained. A maximum relative error value of 8.95% was obtained, which is very well within the range of training quality criteria. The results show that the ANN can profoundly predict the simulation outcomes for a wide range of ionic strength (IS) and can be directly used to obtain the value of dependent variables through simple calculations using network weights and transfer functions.