In the computational molecular sciences, it is often necessary to solve the equations of quantum chemistry or sample a marginal probability distribution of very high dimension. From these calculations, electron densities, energies, and electronic and thermodynamic properties of a system can be extracted, albeit at very high computational cost. The use of machine learning techniques, which are becoming increasingly utilized in theoretical and computational chemistry, can ameliorate these computational costs, replacing them with simple, trainable models. In this talk, I will present the approaches we have been exploring for leveraging machine learning in the generation of orbital-free electronic density functionals1, potential energy surfaces for condensed-phase calculations, and extraction of free-energy landscapes from enhanced sampling calculations2. The performance of different machine learning models, including neural networks, Gaussian kernel-ridge regression models, and least-squares support vector machines will be compared for performance and accuracy as a function of the size of the training data set. Applications to individual molecules, molecular crystals, and oligopeptides will be presented.

Machine learning of electron densities, energies, and free energies: Bypassing the Schrödinger equation and costly phase-space integrals

2019 Spring National Meeting

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