Description
New chemicals and new materials have transformed modern life: pharmaceuticals, pesticides, surfactants, alloys, catalysts, polymers, battery electrodes, and countless other materials play critical roles in healthcare, construction, energy, and other wide-ranging industries. New materials are not generally stumbled upon by happenstance, but rather are discovered through a long process that involves extensive physics-based computer simulations at the atomic level. Electronic structure calculations play an important role in the discovery process, but they can be extremely computationally expensive. As such, there is a long history of approximation methods that trade off speed and accuracy.
Machine learning has the potential to open a new frontier on this speed-accuracy trade-off, and in doing so, significantly accelerate discovery of new materials. In this dissertation, we first cover the quantum mechanical background necessary to understand the problem setting, written with the machine learning community in mind as the audience. Next, we survey the learning-based methods that are pushing the speed-accuracy frontier, along with some foundational non-learning-based methods. Lastly, we investigate self-supervised learning as a mechanism for understanding the shape of the potential energy surface without expensive-to-obtain supervision on energies and forces.