Abstract
Learning and Correcting Static and Dynamic Properties of Materials in Atomistic Simulations
Michele Ceriotti
MLPRC2
Atomic-scale modelling makes it possible to predict from first principles the properties
of materials and molecules. More often than not, however, one has to strike a balance between accuracy and efficiency -- be it in the form of approximate descriptions of the electronic structure or in the form of accelerated sampling protocols that disrupt the natural time evolution of the system. Here I will present two examples of how one can remedy to these shortcomings. First, I will present examples of how Gaussian process regression can bridge the gap between different levels of theory, and even identify the portions of a molecules that are most affected by a given approximation. Second, I will discuss how the dynamical perturcation introduced by (generalized) Langevin sampling, used to improve sampling efficiency, can be predicted and corrected for.
of materials and molecules. More often than not, however, one has to strike a balance between accuracy and efficiency -- be it in the form of approximate descriptions of the electronic structure or in the form of accelerated sampling protocols that disrupt the natural time evolution of the system. Here I will present two examples of how one can remedy to these shortcomings. First, I will present examples of how Gaussian process regression can bridge the gap between different levels of theory, and even identify the portions of a molecules that are most affected by a given approximation. Second, I will discuss how the dynamical perturcation introduced by (generalized) Langevin sampling, used to improve sampling efficiency, can be predicted and corrected for.