(431f) Parameter Sensitivity Analysis of a High-Shear Wet Granulation Model for Experimental Design and Parameter Estimation. | AIChE

(431f) Parameter Sensitivity Analysis of a High-Shear Wet Granulation Model for Experimental Design and Parameter Estimation.


Yeardley, A. S. - Presenter, University of Sheffield
Bellinghausen, S., University of Sheffield
Milton, R., University of Sheffield
Litster, J. D., The University of Sheffield
Brown, S. F., University of Sheffield
A typical workflow in developing a new product via high shear wet granulation (HSWG) involves experimentally testing every combination of input variables, across 3 to 5 scale of operation. When done using traditional DoE approaches, the experimental effort increases as 2n where n is the number of process and formulation parameters. This makes HSWG a costly process to formulate a design of experiments in time, money, and materials. Therefore, this work aims to apply a model-driven design workflow to identify the most critical process parameters and thus reduce and better target experiments to be performed.

The model-driven design focuses on developing a predictive and well-calibrated process model. The process model is based on a population balance modelling framework. Mechanistic understanding of the rate processes is incorporated through appropriate kernels. The most impactful modelling parameters of these kernels must be identified in addition to the operating parameters for the process. To do this, we adopt a Gaussian Process (GP) surrogate modelling approach to directly interrogate the wet granulation computational model. The GP is trained from an input space of the wet granulation computer model produced from randomly sampling the model parameters, normally distributed about the mean. Predicting the model output using a GP enables a reduction in the considerable computational effort required to analytically calculate the Sobol’ indices for a Global Sensitivity Analysis (GSA).

Cross-validation ensures the GP surrogate model is capable of predicting the model outputs given all twenty modelling parameters. Subsequently, calculations of the Sobol’ indices demonstrate that only four of the twenty parameters have sufficient impact to influence the DoE. The most critical process parameter is the liquid spray rate, dominating all four models outputs. The collision coefficient and the critical pore saturation are together the most important modelling parameters with respect to the fine fraction and D50, with each variable ascribing to 20% of the output variance alone and above 60% when including cross-effects. Whereas, for fine fraction the breakage coefficient is the most significant modelling parameter, resulting in a first-order Sobol’ indices of 25% and negligible interactions. Interestingly, the total Sobol’ index of the nuclei-to-drop diameter ratio is 50% making it the most dominant variable for granule porosity but it has little relevance towards the other outputs. Overall, the GSA has identified the critical process parameter and the impactful modelling parameters, and has enabled a proposal of an appropriate experimental design and model calibration workflow. By applying this workflow within model-driven design, industrial development is improved by determining more beneficial production conditions and reducing experimental effort by 30-60% compared to conventional approaches.