Using An AI/ML Based Digital Platform To Build A More Robust And Capable Process

June 28, 2023

Contributed Commentary by Michael Bartlett, 

June 28, 2023 | Legacy methods for developing a robust process and understanding process capability are limiting.  Leveraging modern AI-based digital solutions throughout the product lifecycle provides expansive capability for creating more robust and capable processes.     

To review some terminology – process capability is the measure of the processes’ ability to make a product that meets specifications.  To oversimplify, subject matter expertise and statistics are used to come up with the associated KPIs for process capability.  Process robustness is the ability of a process to adapt to variability in factors like raw materials, process equipment, human error, and others. 

The steps that a process development scientist typically follows are similar from molecule to molecule.  They will screen some media types, process conditions like pH and temperature shifts, filters, columns, etc. with the goal of trying to prove out what they know worked historically, and then refine from there.   

With the ability to digitally model that experiment, and hosts of experiments thereafter, a scientist who typically relies on tribal knowledge to build a recipe can instead rely on the digital tool to become more efficient in recipe development. The move to a digital platform for process development will enhance a scientist’s understanding of the process, remove some confirmation bias (meaning push them to let the data be their guide, as opposed to targeting things that worked in the past), reduce the number of experiments, and help enhance the robustness of the process. 

Digital tools that leverage artificial intelligence (AI) and machine learning (ML) can build digital representations of a process, with a model becoming more accurate with more data.  This model represents a data-based digital twin of a process, and with the identification of critical process parameters (CPPs) and critical quality attributes (CQAs), can be leveraged to fully explore the design space, potentially far outside the boundaries when compared to tools that rely on SPC.  From a process robustness standpoint, the model developed from these exercises can verify if the process can recover past typical upper and lower statistical limits to produce CQAs within spec. In addition to that, the models can be used to explore the full process capability beyond what typical experiments can support.  Using the model to create recipes not typically followed has the potential to improve CQAs like yield. The model can also be trained on different types of starting materials or be subject to simulated process variability and other situations in an effort to gain additional understanding of how that process will react when faced with a non-ideal circumstance.   

A pharmaceutical company that plans to manufacture products they develop may try to test variability in the lab, and account for that in the product development schedule, but a digital representation will allow the scientist to do this more quickly.  In the case of a contract development and manufacturing organization (CDMO), they will not have the time that the sponsor has, so using a digital representation of the process will allow them to go beyond the spaces that they normally explore and dramatically increase their development capacity while improving both the robustness and capability of the process. 

An additional benefit of the AI/ML based digital platform is that models developed in product development can support more efficient tech transfer and scale-up to commercial manufacturing via a mechanism called transfer learning and be a knowledge management system for all aspects of the product lifecycle, including as a real-time optimization tool in manufacturing.  This capability ensures the continuity of the robust process into production, providing the ability adapt to unexpected issues that arise during a batch, predict when the process may be running towards a deviation, as well as recommend setpoints to keep the process trajectory moving towards the most ideal outcome given the situation.   

This is truly the foundation for a Quality by Design approach, where the process is designed to adapt to non-ideal situations and produce a very high percentage of successful outcomes. A manufacturer does not need to start using the digital platform in product development to design a robust process; better process understanding can be achieved by using the same capabilities with manufacturing data.  Using historical data, the AI/ML platform can learn how the process works, understand the CPP to CQA relationships, and provide recommendations for adapting to variability with the goal of achieving not only acceptable quality and performance, but optimal quality and performance given the current process conditions.       

Switching back to process capability, a manufacturer isn’t getting the full story if they are not tracking equipment performance.  Process performance can obviously be impacted by poorly performing equipment.  Whether it is real-time or predicted equipment performance, not knowing the current and potential health of your equipment introduces inherent risk to the success of your process completing within spec, or even completing at all.  Poor performance can introduce inputs to the process that have negative effects, but without the proper system in place, the result may not be realized until the end of the batch.   

For example, in stirred tank bioreactors, we know that agitation is precisely characterized and calibrated to provide the optimum benefit while minimizing stresses to the cells.   Excess vibration from an agitator that is starting to fail can cause an abnormal reaction in the process with the potential result being an underperforming batch.  Even without vibration sensors, data coming from the process, properly modeled, can alert the operator to abnormal performance.   

A robust process aided by AI/ML can predict whether that anomalous behavior in the agitator will cause the batch to underperform, and if the model has seen this situation before, offer up setpoint recommendations to keep it within spec.  By using these digital resources, an operator will know the true process capability, and it will be robust.  If the models say that it cannot be completed within specifications, the operator can take remedial action if it is available, or cut their losses and start over, saving time and materials.   

With advanced digital products purpose built for Life Sciences, the capabilities of all teams in a pharmaceutical manufacturing organization can be advanced considerably, with faster development times and more robust and capable processes. Batch MVDA allows manufacturing operations to understand their process better, and build real-time models for process improvement.  Using AI/ML based process optimization, a manufacturer can even move towards autonomous optimization, making the process even more robust with quick action.  With advanced reliability software the organization can keep the equipment running in the best possible condition.  These capabilities joined together in a single system can keep your operations running at their peak performance, all the time.   


Michael Bartlett is an Account Executive at, focused on helping Life Science customers advance their digital transformation strategies with industry leading technology.  He has 25+ years of experience helping industrial customers accelerate their operational data and analytics strategies through engineering and sales.  He holds a BS in Mechanical Engineering from Northeastern University. He can be reached at