Continuous chromatography AI applications are becoming a practical necessity rather than a convenience, because the complexity that makes continuous purification so efficient is precisely what makes it hard to run by hand. Multi-column periodic counter-current and simulated-moving-bed systems improve productivity, resin utilisation, and facility footprint over batch chromatography, but they do so by running several columns in staggered, interdependent phases at once, generating a decision space and a data volume that manual control cannot manage well. Machine learning is what turns that complexity into something controllable in real time.
This article closes Separation Science's coverage of AI in process chromatography. It builds on the process chromatography guide and connects to the related spokes on AI-assisted Protein A loading, real-time process monitoring, and resin lifetime prediction.
Key Takeaways
|
What Continuous Chromatography Is (and Why It's Complex)
Continuous chromatography replaces the load-wash-elute-regenerate cycle of a single batch column with an integrated system in which multiple columns operate simultaneously in staggered phases, so that feed is processed continuously rather than in discrete batches. The best-established approach for antibody capture is periodic counter-current chromatography (PCC), and the economic case is well documented: PCC improves process economics through higher resin capacity utilisation and lower buffer consumption, with a smaller facility footprint from integrated and continuous operation. Those are exactly the levers that matter most in the cost structure of biologics manufacturing.
Working in analytical science?
Register for a FREE Separation Science account to subscribe to the Separation Science Newsletter.
The reason continuous chromatography is hard is the same reason it is efficient: the columns are interdependent. In a periodic counter-current system, one column is loaded past the point where product begins to break through, and that breakthrough is captured by a second column downstream rather than being lost. This overlapping operation is what achieves the high resin utilisation, but it means the timing of every column switch depends on the state of multiple columns at once, and a change in feed titre or column performance propagates through the whole system.
Aspect | Batch Chromatography | Continuous (PCC / SMB) |
Column operation | One column, sequential steps | Multiple columns, staggered simultaneous phases |
Resin utilisation | Lower; loading stops before breakthrough | Higher; loads past breakthrough, captured downstream |
Buffer and footprint | Higher buffer use, larger footprint | Lower buffer use, smaller footprint |
Control complexity | Manageable manually | Large, interdependent, fast-moving decision space |
Data generated | Per-batch, discrete | Continuous, multi-column, high volume |
Continuous chromatography is efficient for the same reason it is hard: the columns are interdependent. High resin utilisation comes from overlapping operation, and that overlap is exactly what makes the control problem too large to manage well by hand.
Multi-Column Chromatography: How AI Manages Complexity
The core control problem in continuous chromatography is deciding, in real time, when to switch each column between phases as conditions change. This is where AI earns its place, because the complexity of continuous processes makes purely experimental optimisation time-consuming and costly, and model-based approaches have been shown to be superior to experimental trial and error for developing and optimising these processes.
Where machine learning contributes to multi-column management:
- Breakthrough prediction. Models that predict the breakthrough curve of each column from feed and column-state data allow the system to anticipate when a switch is needed rather than reacting after breakthrough has begun, which is central to maximising resin utilisation without losing product.
- Column-switching optimisation. The timing of switches across all columns can be optimised jointly, balancing productivity against resin utilisation and yield, a multi-objective problem that is impractical to solve manually in real time.
- Adaptation to feed variability. When feed titre or composition shifts, as it does in a continuous process fed from perfusion culture, the optimal switching schedule shifts too. Models can adapt the schedule continuously rather than relying on fixed set points.
- Deviation propagation tracking. Because a performance change in one column affects the others, AI monitoring that tracks each column's contribution and flags asymmetries early provides visibility that manual monitoring cannot match at this data density, connecting directly to real-time process monitoring approaches.
The through-line is that continuous chromatography turns process control into a fast, high-dimensional optimisation problem, and that is a class of problem where machine learning, particularly when combined with mechanistic process models, has a genuine and demonstrated advantage over manual operation.
Predictive Control for PCC Systems
The most significant recent advance is making real-time model predictive control of continuous Protein A capture computationally feasible. Model predictive control uses a process model to look ahead and adjust conditions proactively, but the mechanistic models that describe multi-column chromatography accurately have historically been too slow to solve in real time. A 2026 study in Biotechnology and Bioengineering addressed this directly, developing distilled physics-informed neural networks based on the general rate model to accelerate breakthrough curve fitting and four-column periodic counter-current process optimisation, achieving a balance between prediction accuracy and computational speed that makes real-time control practical.
Why physics-informed approaches matter here specifically:
- Speed with mechanistic fidelity. By embedding the governing chromatography equations into the neural network, physics-informed models retain the accuracy of a mechanistic model while solving fast enough for the real-time decisions a running multi-column system requires.
- Real-time parameter estimation. The approach enables process parameters to be estimated live from the running process, so the control model stays calibrated to the actual system state rather than drifting from assumptions made at set-up.
- Proactive rather than reactive control. With a fast, accurate model available, the control system can predict how the process will evolve and adjust column switching and flow proactively, which is the defining advantage of model predictive control over reactive feedback.
This is the same physics-informed modelling approach discussed for AI-assisted Protein A loading, applied at the level of the whole continuous system rather than a single capture step. The convergence is not a coincidence: hybrid physics-plus-ML modelling is emerging as the dominant paradigm across process chromatography because it combines the generalisation of mechanistic models with the speed and adaptability of machine learning.
AI Optimisation of BioSMB and Simulated Moving Bed Workflows
Simulated moving bed (SMB) chromatography and its biopharmaceutical implementations extend the multi-column principle further, using a larger number of columns and more complex switching to approximate a true counter-current process. The optimisation and control challenges scale with that complexity, and the case for AI assistance scales with it too.
The specific optimisation problems in SMB-type workflows:
- Higher-dimensional switching schedules. With more columns and more complex phase relationships, the switching schedule has more degrees of freedom, increasing both the potential efficiency gain and the difficulty of finding the optimum manually.
- Zone and flow-rate optimisation. SMB performance depends on the flow rates in each zone of the system, and optimising these jointly against productivity, purity, and yield objectives is a natural fit for model-based optimisation.
- Robustness to disturbance. The more tightly integrated the system, the more a disturbance propagates, so control strategies that maintain performance under feed and process variability are especially valuable, and AI-assisted control is well suited to maintaining them.
- Integration with upstream perfusion. When a continuous capture system is fed directly from a perfusion bioreactor, the two processes are coupled, and coordinating them, matching capture capacity to a variable feed, adds another layer where predictive optimisation helps.
Across both PCC and SMB-type systems, the pattern is consistent: the more columns and the tighter the integration, the larger the optimisation problem and the greater the advantage of model-based and AI-assisted control over manual operation or fixed set points.
Integrating Continuous Downstream Data With AI
Continuous processing generates continuous data, and the value of AI in this setting depends entirely on integrating that data across columns, sensors, and control systems. This is the practical prerequisite that determines whether the modelling advantages above can be realised at all.
The data integration requirements for AI-enabled continuous chromatography:
- Real-time multi-column data streams. The state of every column, load status, breakthrough, pressure, must flow to the control model continuously and with low latency, because the switching decisions depend on the current state of the whole system.
- Sensor integration under a PAT framework. Inline sensors feeding the control model connect continuous chromatography directly to the process analytical technology approaches used for real-time monitoring, making the two capabilities part of one integrated system.
- Coupling to upstream data. Where the process is fed from perfusion culture, integrating upstream feed data lets the capture system anticipate and adapt to changes in what it will receive, rather than only reacting to them.
- Historian and traceability. A continuous process still produces batches for release, so the data system must preserve the traceability that links product to process conditions across the continuous run, which matters for both quality and regulatory purposes.
The infrastructure lesson is the same one that runs through all of process chromatography AI: the modelling is only as good as the data pipeline feeding it. Continuous processing raises the stakes because the data is continuous and the control decisions are live, which makes robust, low-latency data integration not optional but foundational.
Scale-Up and Regulatory Considerations
Continuous chromatography sits at the intersection of two things regulators are actively engaging with: continuous manufacturing and AI-assisted process control. Both are viewed favourably in principle, continuous manufacturing for its potential quality and consistency benefits, AI-assisted control as an extension of process understanding, but both carry specific expectations that must be addressed.
The considerations that govern moving continuous AI systems into production:
- Model validation for a control role. A model that actively controls a process, not just monitors it, is a higher-stakes application, and it must be validated for that control role with defined performance criteria and behaviour under the range of conditions it will encounter.
- Batch definition in a continuous process. Continuous processing requires a clear definition of what constitutes a batch for release and traceability, and the data and control systems must support that definition rigorously.
- Change control for adaptive control models. A control model that adapts to process conditions raises the question of whether it still behaves like the validated version, which must be managed through change control before adaptive control is used in production.
- Lifecycle governance. The method and model lifecycle framework in ICH Q14 provides a useful governance model for how these process models should be developed, validated, and maintained as the process and the model evolve over their operational life.
The regulatory direction is supportive of both continuous manufacturing and well-validated AI-assisted control, but the rigour required scales with the role the model plays. A model that controls a continuous process carries greater validation and change-control obligations than one that merely advises an operator, and designing for those obligations from the start is what makes the difference between a system that reaches production and one that stalls in development.
What This Means for Your Lab If your organisation is moving toward continuous or intensified downstream processing, AI-assisted control is not an optional enhancement but a practical requirement for managing the complexity that continuous operation introduces. The clearest near-term value is in model predictive control of multi-column capture, where physics-informed approaches have made real-time control computationally feasible, and in the breakthrough prediction and column-switching optimisation that maximise the resin utilisation continuous processing is designed to deliver. Treat data integration as the foundation, since continuous control depends on continuous, low-latency, well-connected data, and design for the model validation and batch-definition requirements that a continuous, AI-controlled process carries from the start. For the surrounding downstream context, the AI in process chromatography guide and the AI in analytical science overview map the full landscape. |
This article was produced under Separation Science’s AI Editorial Guidelines
