Adsorption capacity prediction and optimization of electrospun nanofiber membranes for estrogenic hormone removal using machine learning algorithms

Abstract

This study focuses on developing four machine learning (ML) models (Gaussian process regression (GPR), support vector machine (SVM), decision tree (DT), and ensemble learning tree (ELT)) optimized and hyperparameters tuned via genetic algorithm (GA) and particle swarm optimization (PSO) to analyze and predict the adsorption capacity of four estrogenic hormones. These hormones are a serious cause of fish femininity and various forms of cancer in humans. Their adsorption via electrospun nanofibers offers a sustainable and relatively environmentally friendly solution compared to nanoparticle adsorbents, which require secondary treatment. The intricate task is to find the relationship between input parameters to obtain optimum conditions, which requires an efficient ML model. The GPR integrated GA hybrid model performed the most accurate and precise results with R2 = 0.999 and RMSE = 2.4052e−06, followed by ELT (0.9976 and 4.3458e−17), DT (0.9586 and 2.4673e−16), and SVM (0.7110 and 0.0639). The 2D and 3D partial dependence plots showed temperature, dosage, initial concentration, contact time, and pH as vital adsorption parameters. Additionally, Shapley's analysis further revealed time and dosage as the most sensitive parameters. Finally, a user-friendly graphical user interface (GUI) was developed as a predictor utilizing the optimum hybrid model (GPR-GA), and the results were experimentally validated with a maximum error of < 3.3% for all tests. Thus, the GUI can legitimately work for any desired material with given input conditions to efficiently monitor the removal concentration of all four estrogenic hormones simultaneously at wastewater treatment plants.