Machine learning methods for the simulation of physical systems

In recent years it has been shown that machine learning methods may be used to describe interactions between particles in simulations by constructing so-called inter-atomic potentials. These models are capable of describing interactions with the accuracy of quantum mechanical methods, but without the limitations in system size and run times. There are several challenges in developing such a model predominantly arising from data selection, data representation, and model characterization. Furthermore, it is not yet clear to what extent these models can be  extended into more generalizable and transferable forms. In this project, we aim to develop machine learned inter-atomic potentials for the room temperature ionic liquids, a group of materials which find uses in battery technology, solvents, drug design, and many other areas. The development of a potential for such complex materials requires a more streamlined and automated approach in order to ensure the difficulties mentioned above are adequately addressed. To this end, we will develop a comprehensive framework for the construction of machine learned potentials which can identify unique stages of the process and optimize them accordingly. As a further point of interest, we will study how methods of transfer learning may be applied to systems of ionic liquids, in order to work towards the development of more generalizable models. This work will not only result in the development of state of the art models with which simulations may be performed, but will also streamline the development of future models.