The computational resources needed to model chemical systems of realistic size at a high level of theory is enormous and often times prohibitive. Over the past few years, machine learning (ML) has proven to be a game changer for quantum chemistry, with computational costs decreasing by several orders of magnitude. In this work, we discuss recent applications with a deep neural network (DNN) for molecular property prediction. In one such project, we successfully train a DNN to predict DFT-computed molecular dipole moments over a diverse set of chemical structures and show that the model is extensible to datasets containing large drug molecules and tripeptides (with mean sizes significantly larger than those used for training). Furthermore, we present results showing how a DNN can be used to generate a novel charge partitioning scheme based on training to succesive orders of the electrostatic potential, such as molecular dipole and quadrupole moments. In an effort towards utilizing ML for experimentally-relevant quantities, we will also show entirely ML-generated vibrational (infrared) spectra of drug molecules. The ultimate goal is to convey the fast and accurate predictive power that a DNN can have on complex chemical systems, which is important for numerous applications such as high-level dynamical simulations and materials discovery.

Novel and transferable charge partitioning scheme using a deep neural network

256th National Meeting (2018)

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