(497b) Implementation of Genetic Algorithms In the Generation of High-Order Statistical Models | AIChE

(497b) Implementation of Genetic Algorithms In the Generation of High-Order Statistical Models

Authors 

Domagalski, N. - Presenter, Bristol-Myers Squibb Co.
Hallow, D. - Presenter, Bristol-Myers Squibb Company
Fenster, M. - Presenter, Bristol-Myers Squibb Company
Hobson, L. - Presenter, Bristol-Myers Squibb Company
Tabora, J. - Presenter, Bristol-Myers Squibb Company


To support the quality by design (QbD) philosophy, it is critical to understand the behavior of chemical reactions used for API manufacturing. The parameter response space may be mapped experimentally and further explored by mechanistic or statistical modeling. In this work, the formation of a pharmaceutical intermediate was studied via a series of fraction factorial experimental designs. The results were then used to construct statistical models for several responses, including starting material conversion, desired product formation, and key impurity formation.  The models were constructed using a genetic algorithm to select optimal variables from all the possible linear, collinear and quadratic terms.  Although the models demonstrated good agreement with many of the experimental responses, product formation was not accurately predicted suggesting either experimental error of inadequate representation from a quadratic statistical model.  To improve the models different approaches were explored to increase the order of the statistical model.  The resulting models had much better predictive power than the quadratic models. This talk will discuss the systematic construction of high order statistical models and their implementation in the QbD paradigm.