Integrated dual adaptive control of continuous chromatographic separation processes via reinforcement learning

Area of Science:

• Chemical Engineering
• Process Control
• Artificial Intelligence

Background:

• Continuous chromatographic processes exhibit complex nonlinear dynamics and interdependencies.
• Real-time control is challenging due to cyclic operations and variable interactions.
• Increasing process variability necessitates adaptive control strategies for sustained performance.

Purpose of the Study:

• To develop a reinforcement learning (RL) control framework for adaptive operation of continuous chromatographic systems.
• To enable robust control that maintains performance under dynamic and uncertain operating conditions.
• To improve key performance metrics such as product purity, yield, and operational efficiency.

Main Methods:

• A reinforcement learning (RL) controller was developed and trained using a mechanistic simulation environment.
• The RL agent learned to coordinate multiple process inputs based on observed system states.
• A phase-dependent reward formulation was implemented to balance purity, yield, and efficiency.

Main Results:

• The RL controller successfully maintained target purity despite feed variations, flowrate disturbances, and measurement noise.
• Cyclic steady state (CSS) was achieved within 10 process cycles.
• Process yield improved from 86% to 91%, and product recovery increased by 79% compared to nominal operation.

Conclusions:

• Reinforcement learning provides a robust and adaptive control solution for complex, multivariable chromatographic systems.
• The developed RL framework facilitates intelligent and adaptive process operation.
• This approach offers a pathway to enhance performance and efficiency in continuous chromatographic processes.