Small sample data-driven interpretable artificial neural network computation for two-component chromatographic...

Area of Science:

• Chemical Engineering
• Analytical Chemistry
• Computational Chemistry

Background:

• Chromatographic separation processes rely on models for design and calculation.
• Mechanism-driven models struggle with complex adsorption mechanisms and relationships.
• Accurate modeling is crucial for optimizing separation efficiency.

Purpose of the Study:

• To develop a data-driven artificial neural network (ANN) model for chromatographic separation processes.
• To improve the calculation accuracy for single and two-component competitive adsorption chromatography.
• To create an ANN model that functions effectively with limited sample data.

Main Methods:

• Proposed a data-driven ANN model integrating recurrent neural network (RNN) and plate theory.
• Incorporated physical meanings into ANN layers and used mass conservation constraints to reduce unknown parameters.
• Tested the ANN model against numerical solutions of the Equilibrium Dispersion (ED) chromatography model.

Main Results:

• The ANN model accurately fitted single and two-component competitive elution curves across various adsorption systems (Langmuir, Bi-Langmuir, Sips).
• The model demonstrated effectiveness even when driven by only nine elution curves.
• Achieved an average coefficient of determination (R²) greater than 0.985 for prediction sets.

Conclusions:

• The proposed ANN model offers a robust and accurate solution for chromatographic separation calculations, especially with complex adsorption.
• This data-driven approach overcomes limitations of traditional mechanism-driven models.
• The physically-informed ANN model efficiently handles limited data, paving the way for improved chromatographic process design.