PAT

.

.

Process Analytical Technology (PAT) is by many considered to be a paradigm of recent origin; the official framework targeting the food and pharmaceutical sectors was issued by the FDA in this millennium. However, there are numerous examples in industry and academia, where the underlying principles have been applied for more than half a century. So, wo ist der Witz?

PAT in a nutshell concerns intelligent process monitoring and predictive control aiming at minimising process disturbances, emissions, and “scrap work” (production batches needing a serious workover, before they can be released) while maximising throughput, quality, and most importantly profitability. In other words intelligence built into the process itself.

On a professional corporate level PAT often involves a group of people with different technical skills and competences. The following self-explaining graphic outlines some (if not all) steps involved in setting up a PAT modality in an existing process/plant. All of them are important and can have radical influence on the success/outcome of the PAT project.

The graphic represents my own personal view on a typical PAT implementation. It could be a project/study, where one wishes to optimise a production process, which traditionally has been scarcely instrumented with sensors and/or not reference sampled too often. Hence, little is usually known about the underlying mechanisms regulating the process.

In such case identifying a suitable sensor (PAT modality) capable of either directly or indirectly reading the phenomenon of interest (could be any physical or chemical parameter somehow linked to “product quality”) is of utmost importance. Variation in the sensor signal needs to relate/correlate with the sought variation in the process. This relationship can be a simple linear function (many times it is), but it can also be non-linear; in some cases even non-existing. Non-linearity can be dealt with using suitable data processing and analysis techniques. The best approach is to approach people in either industry or academia, who “have been there, done that” – but not necessarily have published or otherwise communicated (!) the results of their endeavours. Networking and personal relations is everything.

Secondly, retrieval of calibration, validation, and prediction samples (yes, three statistically independent sample sets from the same process are required in principle) has to be organised. This can be both entertaining and demanding, since you often times do not know exactly when it is “time” to sample the process in order to catch the inherent variation. Sampling plans based on sampling theory can be deduced, however, but they may or may not work in your particular case. The extracted reference samples (for all three sub sets) need to entail all possible future variation in the process. Analyse samples chemically/physically, document and save all samples, and develop the multivariate calibration/validation/prediction models.

Thirdly, the brand-new and qualified PAT sensor modality needs to be interfaced to existing process monitoring and control hardware (SCADA system). Standard Operating Procedures for a lot of new routines must be developed and relevant people trained herein. Frequent model validation (of different severity) is recommended to assure that the sensor predicts correctly under all possible circumstances.

Und die Moral von der Geschichte:

There is far more to the concept of process analysis than drawing a straight line between two points and calling it a calibration. Processes are multivariate in nature and succesfull deciphering of their inherent variability requires appropriate sensor technology, multivariate data analysis, and representative sampling. Respect that and you will prevail as a process engineer.

 

Comments are closed.