Physics-informed neural networks to solve lumped kinetic model for chromatography process

Area of Science:

• Chemical Engineering
• Computational Science

Background:

• Chromatography process simulation using numerical methods is computationally intensive and lacks real-time responsiveness.
• Physics-informed neural networks (PINNs) integrate physical laws with neural network structures, offering a balance between accuracy and speed for scientific problem-solving.

Purpose of the Study:

• To investigate the application of PINNs for simulating chromatography processes, specifically focusing on the lumped kinetic model (LKM).
• To develop and optimize a PINN model (LKM-PINN) for accurate and rapid prediction of chromatography breakthrough curves.

Main Methods:

• Designed and optimized a PINN structure tailored for the LKM, including network architecture, training data distribution (emphasizing breakthrough transition), and model complexity.
• Developed an LKM-PINN model comprising four neural networks, 12 layers, and 606 neurons.
• Estimated LKM parameters using breakthrough curves and inferred performance under varied conditions (residence times, concentrations, column sizes).

Main Results:

• The LKM-PINN model achieved comparable accuracy to numerical methods, with a lower average fitting error (0.075 vs. 0.081).
• Significantly improved fitting speed: 160 seconds for LKM-PINN versus 7-72 minutes for numerical methods.
• Further speed enhancement to 30 seconds achieved with random initial guesses for LKM-PINN.

Conclusions:

• The developed LKM-PINN model demonstrates high accuracy and exceptional speed for chromatography simulation.
• The LKM-PINN is suitable for real-time simulations, enabling applications such as digital twins in chromatography.