S117: The development and application of multivariate models to improve process understanding

Multivariate analysis of process data significantly improves one’s ability to resolve the underlying relationships between process parameters. Such analyses distill large quantities of data into meaningful information so that they can be readily interpreted. This, in turn, can be used to improve process understanding and control. Discrimant analysis has enabled us to identify medium characteristics that coincide with higher bioreactor productivity.  Partial Least Squares (PLS) regression has also improved our ability to maximize resin use, while mitigating its associated risks. Furthermore, we have used PLS to develop soft sensors to predict total bioreactor yield. Soft sensor PLS models were improved through the use of a genetic algorithm to minimize the Root Mean Squared Error of Calibration (determined through bootstrapping) by optimizing the subset of parameters used for analysis. The use of multivariate analysis on our process data has made its interpretation more meaningful and thereby improved the understanding of our process.