An artificial neural network model for steady-state contaminant removal during nanofiltration of ground and surface waters under conditions typical of drinking water treatment is derived and validated. Operating conditions such as flux, feed water recovery, and element recovery (surrogate for cross-flow velocity), and feed water quality parameters including pH, total dissolved solids concentration (surrogate for ionic strength), target contaminant concentration, and where possible the diffusion coefficient were used as inputs to predict the ratio of permeate to feed concentration of the target contaminant. Contaminants reported herein include dissolved organic carbon, precursors to total organic halide, four trihalomethanes and nine haloacetic acids containing chlorine,and bromine, hardness, alkalinity, and total dissolved solids. Additionally, source waters from seven different locations and two commercial thin-film composite membranes operating in a wide range of permeate fluxes and feed water recoveries were considered. Deterministic and pseudostochastic simulations showed that artificial neural networks closely predicted permeate concentrations of each one of these organic and inorganic contaminants. Therefore, neural networks can be used to circumvent difficulties associated with formulating and solving the highly non-linear Nernst-Planck equation to calculate solute removal from multi-component solutions at high recovery. Moreover, neural networks can predict the transport of heterogeneous and difficult to characterize water treatment contaminants such as natural organic matter and disinfection by-product precursors, whose physicochemical properties are unknown. Such models can be used to screen membranes prior to conducting expensive large-scale tests as well as in the better design and interpretation of data obtained from site-specific water treatment nanofiltration studies conducted in support of plant design.