 |
Center for Nonlinear Studies |
Machine learning methods are revolutionizing modern chemistry, materials physics, and molecular biology. The standard tools for atomisticsimulations are ab initio quantum mechanics (QM) methods and classical methods such as force fields (FF). Researchers have had to choose between the accuracy and generality of QM and the computational efficiency of FFs. Recent work using machine learning (ML) aims to achieve the best from both QM and FFs for the prediction of properties such as potential energy, atomic forces, atomic charges, and molecular dipoles. A grand challenge is to build an ML model that is not only accurate and fast, but also transferable, so that a single model can accurately describe diverse atomic configurations. The key to transferability is to build a sufficiently broad data set; this constrains the ML model to learn the correct physics, rather than “memorizing” a narrow set of data points. Active learning techniques provide a way to autonomously explore chemical space, allowing one to build a large and diverse data set. In this talk, I will discuss applications of active learning along with new methods for improving the physics of existing machine learning QM property predictors.
Host: David Metiver