(674e) Can a Computer "Learn" Non-Linear Chromatography?: Physics-Based Deep Neural Networks for Chromatographic Separations

The dynamics of chromatographic pulses are represented by hyperbolic partial differential equations along with non-linear equations describing the adsorption equilibrium. In this work we describe a physics-based artificial neural network framework for adsorption and chromatography emulation (PANACHE) that uses deep neural networks constrained by the underlying physics of the process [1,2]. The addition of physics-based constrained allows us to substantially reduce training efforts and also provides excellent description of the pulse propagation in the entire spatio-temporal domain although only initial and boundary data is used for the training itself. The methodology is demonstrated through simulated examples using the four combinations of the generalized Langmuir model, Langmuir (L)-Langmuir, Anti-Langmuir (AL)- Anti-Langmuir, L-AL and AL-L. We further demonstrate that it is possible to use only experimentally measured chromatographic elution curves to train PANACHE, i.e., without requiring explicit measurements of the adsorption isotherm. This method can have applications for systems whose single/competitive isotherms that are complex to be described.

1. Subraveti, S.G., et al., "Can a computer “learn” nonlinear chromatography?: Physics-based deep neural networks for simulation and optimization of chromatographic processes", J. Chromatogr A, in Press

2. Subraveti, S.G., et al., "Physics-based neural networks for simulation and synthesis of cyclic adsorption processes", Ind. Engg. Chem. Res., in Press