(569bn) Kinetic Analysis of an Organic Synthesis Using Mechanistic and Statistical Models

Reaction models are crucial for pharmaceutical process design, optimization, and control. While knowledge-driven models are preferred for their ability to extrapolate to conditions that they have not been trained at, they are not always viable when a thorough knowledge of the process is not available. For complex mechanisms, building mechanistic models can require many exploratory experiments, resulting in lost time and material. On the other hand, response surface models obtained from a classical design-of-experiments (DoE) approach typically take a single time point per experiment and can fail to accurately predict time-resolved data for reactions which may have more typical exponential growth and decay. Prediction of time-resolved profiles is especially important for understanding impurity formation and consumption profiles for intermediate species. The Dynamic response surface model (DRSM)[1] methodology is a generalization to the response surface model that can capture the dynamic characteristics of the process by modeling the output by a basis of shifted Legendre polynomials.

In this work, an alkylation reaction for the synthesis of a pharmaceutical intermediate is studied using data-rich experimentation. DRSM is used to predict kinetic profiles and the results are compared with predictions from a mechanistic model as well as a standard response surface model (RSM) built on reaction-end data points. Prior to fitting the DRSM, an optimization search is performed to find the best critical time and model configuration for each response. The reaction results in a high number of impurities, for some of which the mechanism is unknown. It is observed that DRSM provides excellent fits to the provided data. For some impurities, DRSM is demonstrated to perform better in predicting optimum operating conditions than both the mechanistic model and the RSM.

[1] Klebanov, N. and Georgakis, C., 2016. Industrial & Engineering Chemistry Research, 55(14), pp.4022-4034.