Large datasets for machine learning and reactor response curves for the ring opening polymerization of lactide

Abstract

Machine learning algorithms employ large datasets from which patterns in the underlying information can be ascertained. In the chemical sciences, the ability of machine learning algorithms to increase the effectiveness of molecular simulation has been established and are reputed to have a role in identifying new drug candidates, finding new bulk phases in materials, and other insights. In the case of physical systems, significant labor must be expended to generate a large dataset. Desirable characteristics of a model physical system include 1) output properties that can be reasonably and accurately characterized, 2) a sufficient range and identifiable number degrees of freedom of input variables (e.g. concentration, temperature, time), 3) a level of reasonable reproducibility, 4) and a system for which the time or material costs are not prohibitive. Given our recent work in lactide and carbonate polymerization, we propose this as a candidate physical system for a machine learning application. Polymers can be accurately characterized in terms of their composition, molecular weight, tacticity, endgroups, polydispersity, and rheology with a range of techniques. Computer automation allows us to rapidly iterate through a range of process conditions. This talk will discuss the ring opening polymerization of lactide using a recently developed software platform and a reaction apparatus constructed for the generation of large datasets.