Fractionation of dyes/salts using loose nanofiltration membranes: Insight from machine learning prediction

Wastewater (WW) served as the crucial indicator for sustainable development, human health, and the ecosystem. Nanofiltration (NF) membranes are efficient in contaminants, dye, organics, ionic strength, hardness, and salt treatment from WW. Membranes such as loose NF play crucial roles in dye and salt removal fractionation. This study proposed an insight into machine learning (ML) techniques based on an established experimental laboratory using a loose NF membrane and loose layer surface functionalization of ultrafiltration (UF) membranes with nano-silver-immobilized polydopamine. For this purpose, the obtained data from experimental work were pre-processed and fed into four different computational models viz: hybrid adaptive neuro-fuzzy inference system (ANFIS), robust-linear regression (RLR), support vector regression (SVR), and multi-linear regression (MLR) for the prediction of fractionation of dyes/salts rejection variables (flux (LMH), rejection (%), rejection of dye/salt (%). Based on feature selection, two different model combinations were established, and four statistical evaluation criteria were used to assess the prediction performance. The results justified that SVR-M2 outperformed other models for predicting flux (LMH) and rejection (%) with 95% and 98% accuracy, respectively. Similarly, hybrid ANFIS-M2 proved merit for modelling the rejection of dye/salt (%) with 72% accuracy. The prediction using all the models was found reliable and satisfactory except with the rejection of dye/salt (%), with ranged between marginal and good. The experimentally designed membrane and ML feasibility are excellent examples of fractionating divalent salts and dye.

Automated summary

Wastewater is crucial for sustainable development, human health, and the ecosystem. Nanofiltration membranes are efficient in treating contaminants, dye, organics, ionic strength, hardness, and salt from wastewater. This study used machine learning techniques to predict dye and salt rejection variables using experimental data, and found that the SVR-M2 model outperformed other models in predicting flux and rejection with high accuracy.