Augmenting chemical databases for atomistic machine learning by sampling conformational space

Abstract

Machine learning (ML) has become a standard tool for the exploration of chemical space. Much of the performance of such models depends on the chosen database for a given task. Here, this aspect is investigated for "chemical tasks" including the prediction of hybridization, oxidation, substituent effects, and aromaticity, starting from an initial "restricted" database (iRD). Choosing molecules for augmenting this iRD, including increasing numbers of conformations generated at different temperatures, and retraining the models can improve predictions of the models on the selected "tasks". Addition of a small percentage of conformers (1 % ) obtained at 300 K improves the performance in almost all cases. On the other hand, and in line with previous studies, redundancy and highly deformed structures in the augmentation set compromise prediction quality. Energy and bond distributions were evaluated by means of Kullback-Leibler ($D_{\rm KL}$) and Jensen-Shannon ($D_{\rm JS}$) divergence and Wasserstein distance ($W_{1}$). The findings of this work provide a baseline for the rational augmentation of chemical databases or the creation of synthetic databases.