Machine learning and computational approaches for designing membrane distillation modules

Abstract

Membrane distillation (MD) is a promising emerging water desalination technology. Commercialization of MD modules has been hindered by ineffective heat recovery and temperature polarization effect. Although hollow fiber (HF) membranes provide the highest area-per-module, they are under-investigated compared to flat-sheet membranes due to the interconnection of geometric, thermal, and hydrodynamic parameters in HF MD process. In this work, the parameters impacting HF MD module design are performed based on multiscale and deep neural network (DNN) models. MD experiments are conducted to train and validate the machine learning and multiscale models. The developed models are used either to explain the effects of geometric, thermal, and hydrodynamic parameters on the permeate flux or to predict the flux of a given set of parameters. The results revealed an increase in flux with the flow rate, velocity, and feed temperature. However, it decreased with shell diameter and module length. Compared to the experimental fluxes, flux predictions using multiscale and DNN approaches were within 14% and 1.2%, respectively. The DNN model converged to a mean squared error of 1.21% (R2 = 0.96) within a few minutes and demonstrated its potential as a favorable tool for module design optimization due to its accuracy, speed, and low computational requirements. The present study effectively exhibits the advantages of using machine learning as a next-generation model for fast module design, optimization, and scale-up of MD technology.