We can easily tailor deep learning (DL) architectures to learn the underlying physics of the problems even from simple chemical representations. However, training a deep artificial neural network typically requires large data sets from experimental characterization or even physics-based models, which are dramatically demanding and correspondingly limited in scope. Here, we present a novel active learning sampling method for the training of the DL models to leverage small size data sets. This approach establishes accuracy and speed up in a closed loop of data modeling and molecular design. As a proof of concept, we report capabilities of this method for accurately predicting high refractive index (RI) of organic molecules in a virtual high-throughput screening approach. The data generation protocol requires calculation of electronic and optical properties of molecules using the density functional theory (DFT) and molecular dynamics (MD) simulation. We demonstrate that the proposed approach can reduce the size of training set by %72 to achieve the same accuracy in comparison to the model trained on a larger set. This is a significant speed up in terms of data acquisition (i.e., DFT and MD calculations). We further apply this method to several DL architectures and discuss the involved challenges to extract insights from DL models as a guide for our active learning approach.

Novel, active learning approach for deep learning of chemical data: Extracting more chemical insights by choosing less

2019 Spring National Meeting

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