A reinforcement learning control strategy for continuous multicolumn chromatographic separation processes

Controlling multicolumn continuous chromatographic processes can be challenging due to nonlinear dynamics, online measurement unavailability and column dependencies. Pre-defined operating schedules cannot adapt to changing conditions or disturbances, leading to suboptimal performance. In this work, we present a reinforcement learning (RL) controller for adaptive operation of the twin-column multicolumn countercurrent solvent gradient purification (MCSGP) process. We develop a single controller that optimises the flowrate trajectory, while phase transitions are determined by concentration-based switching criteria. The RL policy is trained using a mechanistic model for the simulation environment. A state representation based on outlet concentrations, operating phase, past actions, and purity requirements provides direct feedback from the process, while a phase-dependent reward encodes the trade-offs between purity, yield, and efficiency. The controller enforces the purity requirement from process start-up, improves product recovery per cycle by 80%, and achieves cyclic steady state (CSS) operation within 10 process cycles. Under variable feed and flowrate disturbances, the controller adapts its operation to maintain purity. These results demonstrate that RL provides an effective framework for adaptive control of continuous chromatographic processes, enabling improved recovery while maintaining product quality under variability and disturbances. • Reinforcement learning control for continuous multicolumn chromatography • Single policy regulating flowrate across all phases of cyclic operation • Purity constraint enforced from process start-up without predefined schedule • Adaptive operation under feed variability and flowrate disturbances • Product recovery per cycle increased by 80% while maintaining purity