A modeling-based comparison study of data-driven and transport models for forward osmosis-nanofiltration hybrid system

Abstract

The conventional transport modeling techniques for FO-hybrid processes are formidably complex and rigorous. Data-driven models may present a better alternative to conventional transport modeling. Therefore, this study develops artificial neural network (ANN) based data-driven models and Speigler-Kedem based transport models which are then compared for Cellulose-triacetate (for forward osmosis (FO)) and Thin-film-composite (for nanofiltration (NF)) based membranes to predict process permeate flux and permeate concentration. Data visualization through correlograms exhibited significant non-linearity in the experimentally generated lab-scale FO-NF datasets. Furthermore, data validation for flux and concentration revealed that ANN algorithm exhibited superior fit with R2 ranging between 0.9 and 0.65 and mean square error (MSE) ranging between 0.095 and 0.0007, when compared to support vector regression (SVR), decision tree (DT), and multiple linear regression (MLR). The optimized ANN-based model showed high R2 values: 95.03 % and 96.08 % for FO and NF datasets, respectively. However, model accuracy is compromised because of limited variation in training datasets, resulting in higher errors (5–12 %) when compared to the transport based model errors (2 %). Despite larger errors, ANN model is competent in predicting the suitable draw solute and responses for the feed side solute. For MgSO4 feed, ANN predicts most effective draw solute in the order of: (NH4)SO4 ≈ Na2SO4 > K2SO4 > MgSO4. Instead, for larger training dataset, data-driven modeling may be equally effective.