PAT: The Value of Better Control

Dec. 29, 2004
Model Predictive Control (MPC) and better sensor technologies will allow pharmaceutical companies to better control their processes. Pavilion Technologies’ Eric Unrau and Michael Tay discuss the challenges and benefits of implementing MPC for pharmaceutical processes.
By Eric Unrau and Michael Tay, Pavilion TechnologiesSelf-reverence, self-knowledge, self-control. These three alone
lead life to sovereign power. -- Lord Tennyson
The FDA Process Analytical Technology (PAT) initiative has at its basis the objective to understand specific pharmaceutical processes more thoroughly using a variety of analysis tools, including not only new sensor technology but also statistical, fundamental and artificial neural network mathematical modeling. It has been widely published in literature that, given appropriate mathematical models of the relationships that govern process behavior, engineers can achieve new insights into their processes and further leverage these models to achieve higher quality, higher efficiency and a process that is in control. The PAT-based projects attempt to apply new sensor and analytic technologies to improve our process knowledge and achieve an intelligently controlled process.We can control processes not within static recipes, but based on a thorough understanding of the achievement of good quality products.The FDA has in multiple forums (ISPE Washington Conference, June 9-10, 2004; ISPE Annual Meeting, October 25, 2004; Pharmaceutical Quality by Design: Improving Emphasis on Manufacturing Science in the 21st Century, August 5, 2004) reinforced this intent with statements that the objective of PAT is to improve process understanding, whereas one result is the potential to reduce regulatory oversight and burden on companies who have demonstrated good practices and a better understanding of their process through proper PAT implementation.Once a process is better understood, the next step is to directly use that understanding as an enhancement to basic level control and directly control quality and the end-product objective. When you control directly to analyzed, measured or understood quality, you are enforcing a good, in-control process —instead of controlling and documenting that the process is maintained within static operating limits.There are two classical technologies to regulate process operations: regulatory control (PID) and model-predictive control. Better analytics enable the usage of more robust control technologies.Regulatory ControlAs analytic technologies are applied to our processes, we can leverage the better understanding that results to achieve better control. Regulatory control, generally PID, is the workhorse of the control technologies used in some form on almost every control loop. In general, control is managed to a target for temperature, level, pressure or flow, and an actuator is adjusted to maintain the process condition at target. This is the simplest model-based controller that includes tuning constants to react to proportional changes in target error, integrated error that avoids long-term offset, and derivative action that responds to the speed of the control response. Given a dynamic process model and a controller objective, IMC (internal model control) tuning rules can be used to calculate an optimal PID tuning. These rules are the basis of most regulatory controller tuning packages.The results of analytic tools enable a pairing of each control target (pressure, temperature or flow) with the best actuator. With an understanding of the relative influence of each actuator on each target objective, called relative gain analysis (RGA), in control terms the mathematical model can be used to design the optimal regulatory control strategy. In many cases the regulatory control loops are straight-forward and intuitive, but for complex systems or challenging single controllers, RGA resulting from an improved multivariable data analysis (PAT application) improves control significantly.
Predictive control is about knowing what is ahead, not what you just hit.

Model-Predictive ControlModel Predictive Control (MPC) was developed to solve control challenges of complex dynamics, multivariate interactions and direct quality control. (To access a sample MPC diagram, click the "Download Now" button at the end of this page.) For example, an MPC solution was very early industrially deployed to solve challenges in on-line optimization projects in hydrocarbon plants. The plant operator agreed that the mathematically calculated optimum was probably the best operating condition, but that he could not get the plant from where it was to the optimum. His challenge was the large number of operating constraints and trip limits that prohibited success. Let us investigate each control challenge solved by MPC individually. Some processes have long, complex and/or mixed dynamic interactions. As an example, most spray dryers are controlled rapidly by adjusting feed rate of concentrate to the dryer. These fast dynamics are significantly different than the slow response required to stably adjust heated drying air. For this reason, most dryers are controlled with regulatory temperature control by adjusting the dryer feed rate (the fast response). Because the desired or ideal control includes both fast (pump speed) and slow (temperature) relationships, MPC is required for the best dryer control. Other challenging dynamic problems occur in very slow processes, where the process responds hours or even days after a control change is made. MPC mathematics includes a long memory of the past process dynamics and is designed to manage a process in cascade to regulatory control. Very slow processes can be managed at an appropriate frequency. As described in both the dryer and the hydrocarbon example above, MPC uses a multivariate mathematical model to calculate and drive to optimal control action. Two, three, or even tens of interacting variables can be controlled in a coordinated fashion. Variables include fully dynamic predictive models of the controlled parameters, such as quality, throughput and efficiency as well as a variety of operating constraints or limits.Another type of variable included in this architecture is feed-forward control variables or process disturbance indicators. A model-predictive projection, or variable trajectory, of what is calculated to affect quality can respond to disturbance variable changes and make corrective controller action before quality is significantly impacted. In the dryer example above, where ambient humidity is measured, the heater can be adjusted as a weather front moves in and moistens the drying air. The powder moisture is held fairly constant. Finally, MPC has enabled direct control of the process objectives: quality, throughput and efficiency. Analytic technologies can enable predictive quality models that can be controlled based on inferential model results and continuously biased to lab quality results as they are measured. Model-based controllers on composition, moisture content, density and pH have all been successfully deployed. Instead of controlling stream flows or temperatures, tighter quality is achieved simply by directly controlling the process objectives. This is a direct utilization of the results of verified mathematical models. MPC can include multiple variables in complex dynamic and static relationships for direct digital control of quality, of inferred laboratory results or on-line analyzers.Where PAT Makes SenseThere has been much discussion on what PAT is and what types of applications can be done under the PAT initiative. But how does a company make an assessment of where to look within its own operations to apply a PAT implementation? And, how can a company best apply a PAT implementation to improve quality and operations performance?Many manufacturers have asked whether a PAT project should be attempted on a new or existing product/process. This depends on a number of factors. The answers to the questions below can serve as a good starting point in determining if a PAT project should be initiated. Many of these questions will need to be analyzed in parallel, not in series. Which factors are more important will change based on the process, product and company objectives.

Is the process well understood? If the process is not fundamentally understood, the first step is to look at offline multivariable data analysis to see whether current available information can provide useful information when analyzed with data analysis tools. Multi-variable data mining and modeling tools are available to provide significant insight and shed light on complex multivariable relationships to real-world problems. Multiple case studies are also available to assist in this assessment. Once the process is well understood, then existing models can be used to develop an online application, either for a sensor or analyzer or for a full advanced process control application.

Are there quality opportunities or variability with current production? When applying an online application, the goal is to reduce variability. In control applications, almost all measurable value comes from the reduced variability provided by the closed-loop control system. The quantifiable financial value from a PAT implementation is clearly from the same area: reduced variability. If a process displays swings in quality, an inability or long time to recover within control specifications once an upset occurs, or different variable targets are achieved upon recurring production cycles, many times these types of problems can be effectively addressed by closed-loop control systems.

If so, is the source of variability quantified or unknown? In processes where there are a number of known or unknown factors affecting quality, a complete multivariable, and likely non-linear, approach will need to be made. In this case, an MPC approach will be the recommended path to achieve the highest quality and reduced variability. For example, the need for a real-time measurement of a key quality parameter, a highly multivariable, non-linear MPC approach might be reduced to the addition of an analyzer to provide real-time quality information online, at-line or inline to the operator who would make appropriate adjustments.

Where is the product in its lifecycle (clinical trials, production, off-patent, etc.)? Most new products have longer lifecycles than an existing product already in production. In addition, the advantage of a new product to “build in” the PAT validation protocol for the initial production design has large financial consequences. However, existing products with strong market positions that can achieve significant profitability improvement by a step change in improved quality (reduced off-spec and regulatory oversight costs) would be good candidates for PAT implementation. In addition, where existing products are being made in similar or identical production lines, the benefits of implementing multiple PAT applications multiply.

In the case of utilizing control models, such as an MPC application, multiple installations become increasingly profitable for the manufacturer, verses the need to purchase and install multiple hardware analyzers for multiple lines. Again, the approach to PAT implementation will change for each process, but an overall view of the end-goal is important from both a value and cost standpoint.

Would a PAT implementation be justified for this product? Most investments in new processes or technologies require a financial analysis to provide the company with a solid cost-benefits analysis. Most PAT projects follow a similar approach. Areas where companies will see significant financial return include:

  • Quality improvements as a result of reduced off-spec product and an ability to operate closer to target specification;
  • production improvement that leads to higher yields and/or more throughput;
  • reduced time-to-market;
  • more efficient production which reduces energy and raw material use per product produced;
  • improved production flexibility through PAT implementation;
  • potential for a reduction in validation costs and compliance; and
  • potential for flexible post-approval continuous improvement activities.

In addition, there are other possible benefits that are process or manufacturer-specific that could add to the ultimate economic value provided by a particular PAT project. However, when evaluating a potential control project, the financial justification is normally the culmination of a number of the previous factors (process understanding, variability, reduced regulatory risk, etc.)

As in any project, the PAT implementation costs are balanced against the lifecycle value provided by that initiative. Many control and MPC applications in a variety of industries have demonstrated a return on investment in months with recurring value year-over-year. Specifically in the case of MPC, manufacturers have documented recurring annual value for as long as ten years.

Conclusion

Process Analytic Technology utilizes improved sensor and mathematical analytics to better understand pharmaceutical processes. Given a better understanding of what impacts quality, reliability and performance, PAT should be used to operate production facilities better. Two classical control technologies are described that offer ways to leverage PAT results directly and achieve a process in control using these “new” principals.

How a PAT implementation is applied can be assessed effectively through a series of requirements ranging from current process understanding to the financial value to a company of a control project. The company’s key objectives will influence where to best apply a PAT implementation in your operations. PAT enables manufacturers to turn data into knowledge and knowledge into action.


About the Authors

Eric Unrau has 10 years of experience in process engineering and business development in the chemical process and consumer products industries. Eric has experience working with process technology in crystallization, various inorganic processes, environmental processing in reduction and elimination of waste streams, and bulk powder and solids processing. He currently serves as Sales Account Manager at Pavilion Technologies, a software company dedicated to helping manufacturers significantly improve plant performance and efficiency. Eric can be reached at [email protected].

Michael Tayis a Technical Account Manager with Pavilion Technologies, where he is involved in evaluation and development of novel industrial applications in optimization and control. With almost 20 years of experience developing and deploying energy and production optimization projects, Michael has focused the last three years on dry- and wet-milling, drying and pharma investigations. He holds an M.Sc. in Chemical Engineering and a B.Sc. in Biochemistry.