Testing functional polyketones to remove scaling calcium and magnesium from real reverse osmosis concentrate and optimizing the process using machine learning models

Abstract

Functional polyketones (FPKs) were synthesized and tested for the adsorb calcium (Ca2+) and magnesium (Mg2+) from real reverse osmosis concentrate (RROC). This study aims to optimize the adsorption process using experimental and machine learning (ML) models, specifically, multilayer perceptron (MLP) and extreme gradient boost (XGBoost). The reaction of the copolymer PK30 with four amines, 1,2-diaminopropane (DAP), 1-(2-aminoethyl) piperazine (AEP), 1-(3-aminopropyl) imidazole (API), and butylamine (BA), produced FPKs. Experimental investigations were conducted to assess the impact of the adsorbent dosage, RROC concentration, and pH on the process. This dataset was used to develop the MLP and XGBoost models after fine-tuning the hyperparameters to forecast Ca2+ and Mg2+ adsorption. These findings indicated that the primary ion adsorption mechanism involved chelation. Higher adsorbent dosages and elevated pH positively influenced the adsorption, although the efficiency notably decreased with increasing RROC concentration. The adsorption efficiencies of the different amine-based adsorbents followed the order DAP > AEP > API > BA, with marginal variations in the API and BA efficiencies. Moreover, SHapley Additive exPlanations (SHAP) revealed the feature importance analysis and their individual impacts on the (MLP and XGBoost) model predictions by providing relationships between the inputs and output. Both MLP and XGBoost models successfully anticipated Ca2+ and Mg2+ adsorption efficiency, establishing a robust input-output relationship with R-2 values of >= 0.94 and >= 0.95, respectively. This study explored Ca2+ and Mg2+ adsorption by FPKs from RROC and introduced ML models capable of optimizing the process with minimal experimental effort.