Adaptable control policies for variable liquid chromatography columns using deep reinforcement learning

Area of Science:

• Biotechnology
• Chemical Engineering
• Process Control

Background:

• Chromatography systems for biotherapeutics face control challenges due to nonlinear behaviors and variable column properties.
• Real-time control is difficult without internal column data, leading to suboptimal performance with static policies.
• Adapting static policies for each column requires costly experimentation.

Purpose of the Study:

• To develop an adaptable, data-driven control policy for chromatography systems.
• To overcome challenges posed by nonlinear dynamics and column variability in bioprocessing.
• To improve the efficiency and productivity of liquid chromatography.

Main Methods:

• Utilized simulation-based data generation and deep reinforcement learning (DRL).
• Developed a controller that manipulates inlet and outlet flow rates to optimize a reward function.
• Trained the DRL controller on a diverse set of chromatography columns with high variability.

Main Results:

• Achieved a single adaptable control policy effective across multiple variable columns.
• Demonstrated higher productivity compared to a human-designed benchmark policy.
• Observed a slight decrease in purity alongside the productivity gains.

Conclusions:

• Deep reinforcement learning provides a promising method for creating adaptable control policies in chromatography.
• The data-driven approach offers a more efficient solution for biotherapeutic downstream processing.
• This strategy addresses the limitations of traditional control methods in dynamic bioprocessing environments.