Algorithms and Tools for Feasibility Analysis and Optimal Experiment Design in Pharmaceutical Manufacturing



The growing need for systematic and risk-based approaches to tackling challenges within the process industry has, by and large, prompted research into applications of various model-based methodologies. Of particular interest to this dissertation is the Quality by Design (QbD) initiative in the pharmaceutical industry. Motivated by their need, this thesis reports several contributions to feasibility analysis and optimal experiment design. Part 1 focuses on feasibility analysis and comprises three chapters which are centered around the adaptation of the nested sampling algorithm for probabilistic design space characterization. Chapter 2 presents the initial adaptation of the nested sampling algorithm, used for general Bayesian computation, for probabilistic design space characterization which is a key activity for practitioners in the pharmaceutical industry. The sampling-based method was shown to be competitive with a recent optimization-based approach, enabling practitioners to exploit the benefits of sampling-based approaches with comparable computational time to other methods. In the following Chapter 3, two strategical and one implementation improvements to the original nested sampling for design space are described, further reducing computational burden thus enabling solution of larger problems. Part 1’s final Chapter 4 focuses on the development of a design centering methodology as an alternative encoding method for sampling methodologies, providing practitioners a format that is convenient to communicate to process operators. Using a surrogate constructed using point samples drawn from the probabilistic design space, the design centering method computes the largest volume box inside the probabilistic design space. Part 2 comprises three chapters containing contributions to optimal experimental design. Partly due to the popularity and effectiveness of sequential experimental design techniques, a signif icant number of recent applications of optimal experiment design seem to forego the concept of experimental efforts. Chapter 5 is presented in attempt to revive the concept of continu ous experimental efforts through showcasing an effort-based methodology to designing dynamic experimental campaigns. Chapters 6 and 7 address two key issues that arises during optimal experiment design when significant levels of model uncertainty are present, a situation often met during the early stages of model development. A bi-objective optimization framework is described in Chapter 6 to mitigate the risk of designing uninformative experiments due to inaccurate initial guesses of model parameter values. The final Chapter 7 is concerned with the design of optimal experimental campaigns in the presence of operational constraints. A i tractable computational framework that integrates the nested sampling algorithm of Part 1 and continuous-effort design approach of Chapter 5 is proposed to tackle the computational complexity. To promote dissemination of the developed methodologies, they are implemented as part of the Python packages DEUS ( and Pydex (