Data-driven forward osmosis model development using multiple linear regression and artificial neural networks

This work investigates the capability of multiple linear regression (MLR) and artificial neural networks (ANN) to model permeate flux in a thermodynamically complex forward osmosis (FO) process. Whey-permeate was concentrated to a dry matter content of more than 55 %, creating a highly supersaturated metastable solution and exceeding the established boundaries of conventional membrane technology. Different ANN architectures were trained and tested with a varying number of hidden layers and neurons to find an accurate structure.
Furthermore, the evaluated significance of the input parameters was used to reduce the model's complexity. This work shows that both approaches (MLR: R²test = 0.9718, ANN: R²test = 0.9849) were able to model the FO's permeate flux accurately, even with a reduced number of inputs. Finally, due to its slightly better performance, the ANN was used to outline the influence of FS inlet flow and process temperature.