Machine learning techniques have clearly proved to be valuable in developing surrogate “easy-to-evaluate” models of molecular properties from experimental or ab initio data. The quality of any data-driven model is intrinsically dependent on the data used to train the model; consequently, well-designed training set can increase the scope of applicability of the model while minimizing the amount of data needed, which can be particularly useful if experiments or ab initio calculations are expensive. We consider this problem in the specific context of developing sparse thermochemistry models of organic molecules using ab initio data. Specifically, given a set of molecules of interest, how to identify a chemically diverse subset of molecules, for which alone ab initio calculations is sought, so that the model trained on this data is applicable to the whole set? To address this, we propose a method that combines a systems approach of design of experiments with a cheminformatics approach for molecule subset selection to systematically and iteratively construct an informative and chemically diverse training data set. We will present this method using illustrative examples involving developing additive models using publicly available large ab initio molecular property data.

Designing an informative training data set for data-driven molecular property prediction

256th National Meeting (2018)

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.