Cascading evolutionary machine learning framework for micropollutants transport behaviors in nanofiltration

Predicting concentration-dependent transport behaviors of organic micropollutants in nanofiltration remains challenging because pollutant rejection arises from complex nonlinear interactions among molecular characteristics, membrane properties and feedwater conditions. Here, we develop a cascading evolutionary machine learning framework that integrates pollutant rejection prediction with concentration sensitivity analysis. The framework accurately predicts nanofiltration rejection performance (R²=0.956) and further enables the extrapolative estimation of concentration sensitivity coefficients for 70 structurally diverse micropollutants. Integration with nanofiltration experiments and molecular dynamics simulations reveal two distinct concentration-dependent transport mechanisms: concentration-enhanced Donnan exclusion governing highly charged contaminants and compression-enhanced permeation associated with large, flexible molecules. These findings establish previously unrecognized links between pollutant structure and concentration-dependent transport behavior. Beyond reducing reliance on extensive empirical screening, the proposed concentration sensitivity coefficient provides a practical indicator for anticipating contaminant breakthrough under fluctuating feed conditions, supporting adaptive operational optimization in membrane-based water treatment systems.