Can a computer "learn" nonlinear chromatography?: Physics-based deep neural networks for simulation and optimization...

Area of Science:

• Chemical Engineering
• Computational Chemistry
• Separation Science

Background:

• Chromatographic process design requires solving complex hyperbolic partial differential equations (PDEs) and nonlinear adsorption isotherms.
• Traditional numerical solvers are computationally expensive, hindering efficient simulation and optimization.
• Accurate modeling is crucial for understanding solute movement in chromatographic columns.

Purpose of the Study:

• To develop a computationally faster and reliable method for simulating and optimizing chromatographic processes.
• To leverage a physics-based artificial neural network framework for emulating adsorption and chromatography.
• To improve the accuracy of process simulations by incorporating physics-constrained loss functions.

Main Methods:

• Utilized a physics-based artificial neural network framework for adsorption and chromatography emulation (PANACHE).
• Developed unique neural network models for each of the four generalized Langmuir isotherm cases.
• Incorporated process optimization routines within the neural network models.

Main Results:

• PANACHE achieved up to 250 times computational speed-up compared to traditional methods.
• Neural network models accurately predicted spatiotemporal concentrations for binary solute mixtures.
• Successfully determined optimal injection volumes for baseline separation of mixture components.

Conclusions:

• The PANACHE framework offers a computationally efficient and accurate approach for chromatographic process design and optimization.
• Physics-constrained neural networks can effectively learn underlying PDEs for improved simulation accuracy.
• This method facilitates precise optimization for achieving desired separations in complex mixtures.