Seminars & Colloquia Calendar

In the last decade, machine learning methods changed substantially the way in which interatomic potentials are constructed from first principles quantum mechanics. In these approaches deep neural networks, trained on electronic structure data, are used to represent the potential energy surface. Molecular dynamics with machine learned potentials has computational cost and scaling with size comparable to those of empirical force fields, yet it retains the accuracy and generality of the adopted ground-state electronic solver. The scheme can be extended to model how the electric polarization in insulators depends on the atomic configuration, making possible to study the evolution of the dielectric properties of materials along atomistic trajectories.  

 I will present three examples of application of this methodology, all of which are well beyond the reach of standard first-principles molecular dynamics methods. In one, the homogeneous nucleation rate of ice from supercooled water was calculated and found to be in good agreement with experiment. In another, the static dielectric constant of liquid water was extracted from the dipolar correlations using both periodic and reaction field (Kirkwood-Froelich) boundary conditions. In the third example, the ferroelectric phase transition of lead titanate was studied, finding good agreement with experiment for the calculated enthalpy, the spontaneous polarization, the specific heat and the dielectric susceptibility.      

Finally, I will comment on current limitations and challenges.