Closing The Gap Between Modeling And GMP Predictive Control

By Gaurav Walia, PQE Group.

Biopharmaceutical manufacturers have invested heavily in process automation, analytical technologies, and digital modeling. Modern facilities can collect extensive process data, monitor critical parameters, and increasingly use predictive models to anticipate changes in manufacturing conditions. Yet the ability to predict what may happen inside a bioreactor has not consistently translated into the ability to intervene automatically.

That distinction is increasingly important as manufacturers explore more advanced process control strategies. For example, predictive analytics can help identify emerging process variability, but predictive control requires more: a reliable connection between the model, the manufacturing equipment, and a control strategy that can act on the model's recommendations.

In a GMP environment, that connection raises questions that extend well beyond technical feasibility. Manufacturers must establish that the underlying data is reliable, the models remain appropriate for their intended use, automated interventions operate within established process boundaries, and the resulting decisions can be reconstructed and defended. These are prerequisites for a predictive model to demonstrate sufficient reliability to influence, and eventually execute, decisions that affect product quality.

Predicting A Process Is Not The Same As Controlling It

Many bioprocessing operations already employ advanced monitoring capabilities. Process analytical technology (PAT), multivariate data analysis, and soft sensors can provide information about process conditions that would otherwise be difficult to measure continuously.

For example, a soft sensor might estimate a difficult-to-measure process attribute using readily available measurements such as dissolved oxygen, pH, temperature, gas exchange, and nutrient consumption. A predictive model could use those inputs to anticipate changes in cell metabolism or emerging deviations from expected process performance.

That information can be valuable even without automated action. Operators may use the prediction to investigate process conditions, increase monitoring, or evaluate whether an adjustment is warranted.

Predictive control introduces a different responsibility. Rather than simply identifying a potential change, the system must determine an appropriate corrective action and, within its authorized operating boundaries, implement that action through the manufacturing control system.

Consider a fed-batch bioreactor in which a model predicts that the current nutrient feed strategy may lead to undesirable metabolite accumulation. A monitoring application might alert an operator to the potential problem. A decision-support system could recommend a modified feed rate. A model predictive controller could evaluate possible feed adjustments against defined process constraints and automatically select an appropriate change.

Each approach depends on predictive information, but inaccurate predictions can have very different consequences.

The more authority a model has over the manufacturing process, the more important it becomes to demonstrate that the model behaves reliably under both expected and abnormal operating conditions.

Why Successful Models Do Not Always Translate Into GMP Operations

Predictive models are frequently developed using historical process data, laboratory experiments, or engineering studies. Their performance may be convincing in those environments, especially when the data are well characterized and operating conditions are relatively stable.

Commercial manufacturing introduces additional variability.

Differences in raw materials, equipment characteristics, sensor performance, manufacturing scale, and process history can affect how model inputs relate to actual process behavior. Biological systems also add complexity because cellular responses may change during a production run. A model developed under one set of process conditions may lose predictive accuracy when those conditions change.

This becomes particularly important when manufacturers move from development or pilot-scale operations into commercial production. A model that accurately predicts process behavior during a controlled development study may require additional verification before it can support decisions in a manufacturing environment.

The issue is not necessarily that the original model was poorly developed. Rather, its intended use has changed, and the evidence supporting that use must be reconsidered.

Manufacturers therefore need to evaluate predictive performance across the operating conditions the model is expected to encounter. This includes understanding where predictions become less reliable, which variables most influence model performance, and how the system should respond when incoming data fall outside the model's established domain.

A useful predictive model should generate an estimate, and it should also provide a basis for recognizing when that estimate should no longer be trusted.

Digital Twins And Digital Shadows Have Different Operational Roles

Digital twins are frequently discussed as an enabling technology for autonomous bioprocessing. However, the term can encompass substantially different implementations, and not every digital representation of a manufacturing process can support closed-loop control.

A digital shadow may receive process information from manufacturing equipment and use it to represent or analyze current operating conditions. Information generally flows from the physical process into the digital environment without the digital system directly controlling the equipment.

A digital twin can support a more interactive relationship, using process data to update a virtual representation and potentially evaluate alternative operating strategies. Depending on its architecture and integration, it may also support decisions that influence the physical process.

A digital shadow that provides accurate process visibility may be highly valuable for monitoring and troubleshooting. A digital twin that can simulate alternative control actions may support process optimization. Neither capability alone establishes that the system is suitable for autonomous GMP operation.

That requires additional evidence on model accuracy, integration with manufacturing controls, response times, cybersecurity, data integrity, and handling of unexpected operating conditions. Manufacturers should therefore evaluate digital technologies based on the decisions they are intended to support.

Validation Must Address The Decisions A Model Can Make

Traditional automation systems typically operate according to defined control logic and established operating parameters. Predictive control systems may introduce models whose outputs depend on complex relationships among multiple process variables. This creates additional considerations for validation.

A manufacturer must establish that the model suits its intended purpose, but that assessment cannot end with statistical measures of predictive accuracy. It must also address how model outputs influence manufacturing decisions.

For example, a model might demonstrate strong average predictive performance while occasionally producing significant errors under less common process conditions. Those errors may be acceptable for a monitoring application that prompts further investigation. They could be unacceptable if the same model independently adjusts a critical process parameter.

The validation approach should therefore reflect the potential impact of the model's intended use on product quality and patient safety.

Important considerations include the quality and representativeness of training data, model performance under relevant operating conditions, the reliability of input measurements, the limits of authorized control actions, and the system's response when model performance deteriorates.

Manufacturers must also determine how model changes will be managed throughout the system's lifecycle.

Some models may remain fixed after qualification, with changes introduced through formal evaluation and change control. Others may be designed for periodic recalibration or retraining. Models that adapt during operation add complexity because their behavior may evolve after initial validation.

In each case, the manufacturer needs a defined approach for evaluating whether changes affect the model's suitability for its intended GMP use.

The FDA has recognized these issues in its discussion of artificial intelligence in pharmaceutical manufacturing, including questions concerning model validation, lifecycle management, data governance, and regulatory oversight.

The underlying principle is familiar to GMP manufacturers: the evidence required to support a system should be proportionate to the risk associated with its intended use.

Human Oversight Should Be Designed Into The Control Strategy

The transition toward greater manufacturing autonomy does not require an immediate shift from manual decisions to fully automated operations. A more practical approach is to expand the authority of predictive systems incrementally as manufacturers establish evidence of their reliability.

Initially, a predictive model may operate in a monitoring capacity, generating forecasts that operators compare with actual process behavior. This allows the manufacturer to assess performance without permitting the model to influence equipment operation.

The next stage may involve decision support, where the model recommends specific adjustments, but an authorized operator reviews and approves the action.

Subsequent implementations may permit automatic adjustments within predefined limits, while maintaining operator oversight and requiring intervention when the system encounters conditions outside its validated operating range.

For example, a controller might be authorized to adjust a nutrient feed rate within established boundaries while being prohibited from exceeding specified limits or changing other critical process parameters. Such an arrangement would allow manufacturers to gain experience with automated control while retaining clear restrictions on the system's authority.

The effectiveness of this approach depends on more than maintaining a human approval step. Operators must understand the information the system presents, the significance of its recommendations, and when intervention is necessary.

Poorly designed human oversight can create its own risks, particularly if operators become accustomed to accepting recommendations without sufficient evaluation. Consequently, manufacturers should consider how decision-making responsibilities are distributed among the predictive system, the automation platform, manufacturing personnel, and quality functions.

Model Performance Must Be Monitored After Deployment

Validation establishes an initial basis for using a predictive model, but it does not guarantee that the model will remain suitable throughout commercial operation. Sensor drift, changes in raw material characteristics, equipment maintenance, process modifications, and shifts in operating conditions can affect predictive performance over time.

Manufacturers need mechanisms to identify when those changes begin to influence model behavior.

This may involve:

• comparing model predictions with subsequent laboratory measurements,
• monitoring residual errors,
• evaluating changes in input data distributions, and
• periodically reassessing model performance against established acceptance criteria.

The control strategy should also specify what happens when performance falls below acceptable limits.

Depending on the application, the system might suspend model-driven adjustments, revert to a conventional control strategy, require operator approval, or initiate a defined investigation.

These fallback arrangements are particularly important for closed-loop applications. A predictive controller should not continue making adjustments indefinitely when the information supporting its decisions has become unreliable. Recognizing degraded model performance and returning the process to an established control state is essential to responsible automation.

The Path Forward Is Incremental, But It Must Be Deliberate

Predictive control offers opportunities to improve process consistency, reduce variability, and respond to changing manufacturing conditions before they cause deviations. Those benefits are particularly relevant in bioprocessing, where process disturbances may have consequences that become apparent only later in production.

However, manufacturers should resist treating full autonomy as the immediate measure of success.

A predictive system that reliably identifies emerging variability and enables earlier operator intervention can provide meaningful value. So can a controller that automates a narrowly defined adjustment within established operating limits.

These applications can also generate the operational evidence needed to support more advanced implementations. The most productive starting point is often a specific manufacturing decision where the process is well understood, relevant measurements are available, and the consequences of an incorrect action can be controlled.

From there, manufacturers can establish the model's intended use, evaluate its performance, define the limits of its authority and demonstrate that the integrated system operates reliably. That progression requires collaboration among process scientists, automation engineers, manufacturing personnel, validation specialists, and quality teams.

Predictive control cannot be implemented successfully as an isolated data science initiative because the model ultimately becomes part of a regulated manufacturing process.

The long-term opportunity is substantial. More sophisticated models, improved analytical technologies, and increasingly-integrated manufacturing systems may allow biopharmaceutical facilities to automate decisions that currently depend on human interpretation. The technology is advancing quickly; however, unqualified progress is not enough. It will be important for manufacturers to demonstrate those technologies can be trusted to make consequential process decisions and that they can maintain the required state of control throughout manufacturing.

About The Author:

Gaurav Walia is vice president of CSV/CSA/DI and digital governance, head of the PQE Chicago office, and a local equity partner at PQE Group. His career of nearly 30 years includes experience in laboratory management, technology implementation, software design, and international regulatory interaction. He is an adjunct lecturer at Temple University and a frequent speaker on quality, compliance, and regulatory trends. He has more than 100 FDA audits/inspections, as well as three major consent decree projects, under his belt. He is also a former director of R&D, quality, and IT CSV. He is a member of PDA and ISPE and a contributing expert with publications for KENX, an executive advisory board member for IVT/Informa, and the vice chairman in Chicago for the American Society for Quality (ASQ).