سيداميرحسين زاهدي قدوسي

The forward osmosis (FO) process has gained significant attention as an emerging technology for water purification an‎d wastewater recovery. Optimizing the membranes used in this process plays a key role in enhancing efficiency an‎d reducing operational costs. Therefore, this study aims to develop a machine learning-based model to predict water flux an‎d reverse salt flux across cellulose acetate/cellulose triacetate membranes. To do so, several machine learning algorithms including multilayer perceptron (MLP), support vector machine (SVM), an‎d ran‎dom forest (RF) models were trained utilizing experimental datasets gathered from relevant literatures. Various membrane synthesis parameters, including the composition of casting solution, coagulation bath conditions during phase inversion, an‎d various operatiing conditions, were considered as 26 independent input features for training the mentioned models. As a result, the MLP models with two hidden layers, Bayesian regularization algorithm as training function, an‎d the tangent-sigmoid activation function, performed exceptionally well (R2‎> 0.99) in fitting the experimental data corresponding with water flux an‎d reverse solute flux. Subsequently, the effect of some parameters on water flux an‎d reverse salt flux was investigated using the resulting MLP architectures as the optimal model. To identify the best trade-off, the optimal MLP models were further introduced to genetic algorithm optimization technique. The optimization convergence after 70 iterations yielded an optimal condition with Jw ≈ 7.08 L.m-2.h-1 an‎d Js ≈ 6.28 g.m-2.h-1. These findings highlight the potential of hybrid machine learning–optimization frameworks to accelerate the FO membrane design by quantitatively balancing permeability an‎d selec‎tivity.

اسمز مستقيم , يادگيري ماشين , غشا سلولز استات , الگوريتم ژنتيك

forward osmosis , Cellulose acetate membrane , machine learning , genetic algorithm

seyed amirhossein zahedi ghoddoosi

toraj mohammadi