(144a) Producing Predictive Chemical Process Models from a State of Incomplete Mechanistic Knowledge and Small Data Sets

ABSTRACT: In this work, we explore the possibility of developing a quantitative and predictive model for chemical processes, without having to first gain a complete understanding of the underlying mechanisms. We find that this can be done efficiently combining: (1) expectations for the functional expression of the process—extracted from a general, qualitative understanding on how the chemical process ought to behave—and, (2) small data sets providing examples of the process outputs resulting from varying process inputs. The strategy discussed is labeled knowledge-constrained machine learning. The main example we take is on reaction kinetic models—and, here, the expectations are written in terms of constraints on the model predictions over time, and the data is a series of time-point measurements for the reaction outcome under different conditions. But, the intent of the presentation is to discuss a general methodology that may be used to develop predictive models across the range of unit operations found in the processes used to produce synthetic drug substance API’s.