The PAT Initiative has always aimed to change the mindset of the pharmaceutical industry to replace ‘test after each manufacturing step to find out what was made’ to 'test and correct during each manufacturing step to ensure that what is made is only what is wanted'. The drugs delivered to the consumer will be just as safe as they are today but with much less effort and cost for the manufacturer. This is the result the Food and Drug Administration (FDA) wants to see.

*But why isn’t this done already?*It was simply not possible, because the process technology solution to the technical problem at the root of it all didn’t exist. The problem is that manufacturers need to be able to predict what the properties of the product at the end of the manufacturing step will be if they continue to operate in their present manner. They need to be able to perform these predictions from analysing the many other measurements that can be extracted during the manufacturing step and use their predicted values to alter some of the step variables so that the final result prediction can be corrected, if required, before an actual deviation occurs. Think about this. Not only have they got to predict the consequences of how they are operating, which implies some kind of mathematical model, but they also have to do it before the thing they are predicting shows any change. This means an on-line analyser will not be much help. Following this they have to work out which of the many possible process variables to alter and by how much in order to correct the impending deviation.

First-principles models which formulate and solve the basic equations of chemical kinetics, equations-of state, thermodynamics and mass balances are not generally available for the time dependent complex chemistry of AI manufacture or the complex material rearrangement processes that occur in a tablet press. Statistical and chemo-metric models need considerable mathematical skills to create and maintain and do not reproduce the non-linear effects common in processes. Neural Net models when built with two hidden layers can reproduce non-linear effects, but then become very difficult to train without over training and so require constant expert support.

Geometric Process Control (GPC) technology, based on multidimensional geometry, offers a single solution to most of these difficulties and requires no mathematical knowledge to implement or maintain. Its models are the geometric envelopes of multi-dimensional operating points from past process history where the desired result was achieved. The process operating objective is converted to a geometric intention of being an interior point in the envelope. The usable space inside the envelope is easily found from the present operating points and shown to the process operator in an easily understandable visualisation in terms only of the existing process and predicted quality variables, as can be seen.

With this new process technology, violations of the envelope are immediately apparent and corrective advice easily and automatically generated from geometry. The envelope model has a dual response to time based events. This allows it to give very early indication with avoidance advice for potential deviations.