Developing physics-informed neural networks for model predictive control of periodic counter-current chromatography

Area of Science:

• Biopharmaceutical manufacturing
• Process modeling and simulation
• Artificial intelligence in chemical engineering

Background:

• Continuous manufacturing in biopharmaceuticals demands sophisticated design, monitoring, and control.
• Traditional mechanistic models are computationally intensive, hindering real-time applications.
• This limitation impacts optimization and control in biomanufacturing.

Purpose of the Study:

• To develop a computationally efficient modeling approach for biopharmaceutical continuous manufacturing.
• To address the limitations of traditional numerical methods in real-time process control.
• To enable advanced applications like digital twins in biomanufacturing.

Main Methods:

• Implementation of a Physics-Informed Neural Network (PINN) integrated with a General Rate Model (GRM).
• PINN-based GRM for accurate and rapid simulation of chromatographic processes.
• Application to offline simulation of breakthrough curves and online optimization of four-column periodic counter-current chromatography (4C-PCC).

Main Results:

• Significant reduction in simulation fitting time: from 2608.6s to 110.7s for offline simulations.
• Successful online simulation of 4C-PCC within 12-14 seconds.
• PINN demonstrates broad applicability across various parameters and effective parameter estimation.

Conclusions:

• The PINN-based GRM offers a highly accurate and reliable alternative to traditional models.
• This approach significantly reduces computational burden for biomanufacturing simulations.
• PINN shows strong potential for real-time model predictive control and digital twin development in bioprocesses.