A Parallel Hybrid Model for Integrating Protein Adsorption Models with Deep Neural Networks

A high-accuracy modeling of the mass front evolution in a fixed bed is determined by considering the equilibrium information of the components’ concentration adsorbed on the stationary phase. Thus, adsorption isotherms based on thermodynamic principles are required in the mass balance. It is challenging to include isotherms with a high level of detail due to the necessity to compile it on each instant and position, drastically increasing the computation time. Pattern identification methods may be a solution to use a high-level isotherms model quickly and efficiently. Here, we structure a Deep Neural Network aiming to train a surrogate model using the solution of a non-linear Poisson-Boltzmann equation as a data set, modified to include the ionic dispersion potential from the Lifshitz Theory. Thus, the surrogate model output is used in the mass balance model while solving the partial differential equations which describe a column mass front. This approach generates a hybrid model in a serial identification strategy. For a case study, we analyze the mass front behavior of lysozyme in a silica-packed-bed column from pH 6 to pH 10, using different salts: NaCl, NaBr, and NaI. The results indicate a low retention time for the pHs near the isoelectric points for lysozyme (pH 10.8) and silica (pH 5.7) [1]. In addition, differences in salt types allow a significant decay in retention time and an increase in mass front compressibility in the following order: I– > Br– > Cl– due to the differences in anion polarizability. The results here indicate that a link between different scales is successfully achieved using DNN.