Development of a knowledge based hybrid neural network (KBHNN) for studying the effect of diafiltration during ultrafiltration of whey

The membrane surface dynamics is very difficult to predict and can be roughly estimated by the available models but a true depiction is always difficult since the magnitude and direction of driving forces change as a function of time. The present study is an effort to address the issue, so that the combinatorial approach of deterministic and stochastic modelling might present a better understanding of membrane dynamics. The effect of diafiltration has also been incorporated to investigate the effects it has on the membrane. A stochastic model developed by a knowledge based hybrid neural network (KBHNN) was trained using the Levenberg–Marqurt algorithm where the film layer model was used as the deterministic layer, called the first principle model (FPM). Present work employs two different types of KBHNN architecture with an effort to understand the suitability and applicability of the hybrid network in case of predictions for an ultrafiltration (UF) process. In one sort of architecture neural part was in series with the FPM and in the other one it was in parallel with the FPM. The high correlation coefficient (R2) value portrays the correctness and preciseness of the underlining assumptions and establishes the validity of the developed network.

Research Highlights
► Primarily the objective of the study was to understand the effect of diafiltration on a membrane separation technique for the recovery of whey protein from casein whey.
► Secondly development of a mathematical model for the process to predict the percentage purity of the whey protein at each stage of diafiltration for each step of the membrane separation.
► Mathematical model developed in the present study was based on the artificial neural network system in conjunction with the mass balance model related to membrane separation. This combinatorial approach for mathematical formulation is called a knowledge based hybrid neural network (KBHNN).
► The present study focuses on a comparative analysis of two types of KBHNN architecture and its merits or demerits over a simple neural network.
► Finally a conclusive idea on the applicability of KBHNN in the membrane separation process.