Model-predictive design includes a hierarchy of different approaches, explains Bernhardt Trout, PhD, director of the Novartis–Massachusetts Institute of Technology (MIT) Center for Continuous Manufacturing (CCM) and professor in the Department of Chemical Engineering at MIT. The most basic approach is the use of linear univariate models (e.g., parameterized from a linear design of experiments). A more complex approach is multivariate analysis, followed by mechanistic modeling; the ultimate approach is analysis of first principles (e.g., conservation of mass and energy).
The pharmaceutical industry needs to move beyond statistical models, says Fernando Muzzio, PhD, director of the Engineering Research Center for Structured Organic Particulate Systems (C-SOPS) and professor in the Department of Chemical and Biochemical Engineering at Rutgers University. "Models based on first principles allow you to achieve a greater level of understanding and thus extrapolate outside of the area defined by the statistical models," explains Muzzio, who further discusses modeling tools in an article in this issue (see "Model-Predictive Design, Control, and Optimization").Generating the level of knowledge necessary to build a model, whether it be statistically or mechanistically based, helps to speed development and mitigate risks during product development, scale-up, and commercial manufacture, adds Howard Stamato, associate director in the Portfolio Enabling Technology Group of Drug Product Science and Technology at Bristol-Myers Squibb. He adds that in those cases where the model can be used in closed-loop control, there are significant advantages for both efficiency and quality of the final product.
Much progress has already been made in employing models in the pharmaceutical industry, driven in part by advances in process analytical technology (PAT). "Advances in computing power and analytical methods, including PAT, have broadened what can feasibly be measured and subsequently modeled," says Stamato. "These advances allow models to be built with reasonable resource commitments and much higher accuracy than previously possible," he adds. In the solid-dosage drug-product manufacturing arena, modeling work has been performed, for example, in scale-up of fluidized-bed processing using computational fluid dynamics (CFD) (1) and in tablet coating using mechanistic models (2) and discrete element modeling (DEM) (3).
Models for continuous, plant-wide solid-dosage processes have also been developed (4–5). The Novartis–MIT CCM, for example, constructed a functioning prototype of an end-to-end continuous manufacturing line that forms coated tablets directly from API and, in late 2012, finished integrating the process with a control system that used models to simulate the entire process. Models for the individual unit operations are described by first principles, such as material balances, energy balances, and chemical kinetics, notes Richard Braatz, the Edwin R. Gilliland Professor of Chemical Engineering at MIT. "An advantage of first-principles models is that you can build a predictive model with less data. Statistical models tend to have poor extrapolation outside of the data range," explains Braatz. MIT used modeling software to connect the individual units and then tuned the process controllers by running simulations of the process as a whole. "Without a model of the process, we would not have been able to complete the project in an appropriate time because it would have taken too much trial and error to tune the controllers," says Braatz, who notes that when starting up a continuous process it is desirable to quickly reach steady-state operations to avoid producing out-of-specification material.