Designing new materials needs a detailed understanding of the structure and processes of matter on the atomistic scale, because all macroscopic properties ultimately depend on microscopic interactions. For such studies, quantum mechanical modelling combined with atomistic simulations has been proven to be predictive in addition to being able to explain experimental phenomena. However, larger length and timescales are not easily accessible due to the non-linear growth in computational resources required to numerically solve the quantum mechanical equations. We would like to enable fast simulations without a compromise in accuracy by using machine learning techniques to fit the quantum mechanical model. To realise this aim, we have developed the Gaussian Approximation Potentials framework, which uses microscopic data from quantum mechanical calculations on small systems to create fast, accurate and scalable models. Apart from data, the other main ingredient needed to fit Gaussian Processes are kernels. In my talk I will discuss kernels that are designed to compare atomic structures and show examples from molecular and condensed matter systems. These kernels are used to define a set of interatomic potentials or models, and a Bayesian approach determines which is the most likely, based on the data as evidence.
Back to Long Programs