Understanding Single-Protein Fouling in Micro- and Ultrafiltration Systems via Machine-Learning-Based Models

Protein fouling is a complex mechanism. To enhance understanding on protein fouling of membranes, the target of this study is twofold: (i) to determine the relative influences of parameters via the Random Forest (RF) model and (ii) to evaluate the predictive capability of the Neural Network (NN) model. Membrane pore size is the most dominant influence on fouling followed by transmembrane pressure (TMP), while membrane configuration (i.e., flat-sheet, hollow fiber, or tubular) is the least dominant. The NN model gives modest predictive capability despite inconsistencies and variabilities of the experimental setups and protocols, which invariably affects the important parameters in the database compiled from past publications. The database was divided into microfiltration (MF) and ultrafiltration (UF) subsets based on the membrane pore size values. It was found that the dominant parameters for permeate flux are different, with membrane pore size and protein concentration being dominant for MF and UF, respectively, while TMP is dominant for protein rejection for both cases. For permeate flux, membrane material is the most dominant parameter for the non-BSA database, while membrane pore size remains the most dominant parameter for protein rejection regardless of the protein used. Results show that such data-driven RF and NN models can enhance the understanding on the relative dominance of the parameters on different phenomena and provide adequate prediction of protein fouling, in the absence of any governing equations.