Integrated models for prediction and global factors sensitivity analysis of ultrafiltration (UF) membrane fouling: statistics and machine learning approach

In this work, machine learning was employed to quantitatively describe nonlinear ultrafiltration membrane fouling behaviors, from existing data process modeling, process analysis and predictive modeling of unknown data prediction and feature analysis. Instead of using the secondary data calculation that the traditional model required, direct observation data is used for modeling and analysis. This simplifies the prediction process and lowers the cost of the prediction. Besides, the problem that the long-term serial data and the short-term rapid change process were difficult to quantify has been resolved. Two distinct prediction models were established: one is semi-automatic prediction of future data with existing data based on statistics (for long-term study and pre-diction), and the other is fully autonomous prediction based on tree model (for short-term). 2520 12-dimentional laboratory measurements were collected enabling precise modeling prediction (less than 4 % error) for 50 % of the future timeline through supervised learning of process modeling. Results revealed that the UF membrane had a strong rejection impact when it came into contact with a polluted environment, which caused an inconsistent self-pollution coefficient and rapid fouling at initially. For process analysis, a global variable-based weighting factor sensitivity analysis and a statistically significant likelihood estimation were conducted using random put-back samples to accurately predict membrane fouling in an uncertain environment (MSE = 0.2 to 0.26). A high-dimensional variable-specific real-time weighting analysis was derived for inform lifespan extension of the UF membrane at environmental relevant conditions. Overall, this study illustrates the feasibility and interpretability of machine learning-based data-driven approaches in quantitatively describing and understanding nonlinear complex dynamics in membrane fouling.

Automated summary

In this work, machine learning was used to quantitatively describe nonlinear ultrafiltration membrane fouling behaviors. Direct observation data was used for modeling and analysis, simplifying the prediction process and reducing its cost. Two distinct prediction models were established: one for long-term study and prediction using existing data, and the other for short-term prediction based on a tree model. The study demonstrates the feasibility and interpretability of machine learning-based approaches in understanding membrane fouling.