Neural network interatomic potential for laser-excited materials

Data-driven interatomic potentials based on machine-learning approaches have been increasingly used to perform large-scale, first-principles quality simulations of materials in the electronic ground state. However, they are not able to describe situations in which the electrons are excited, like in the case of material processing by means of femtosecond laser irradiation or ion bombardment. In this work, we propose a neural network interatomic potential with an explicit dependency on the electronic temperature. Taking silicon as an example, we demonstrate its capability of reproducing important physical properties with first-principles accuracy and use it to simulate laser-induced surface modifications on a thin film at time and length scales that are impossible to reach with first-principles approaches like density functional theory. The method is general and can be applied not only to other laser-excited materials but also to condensed and liquid matter under non-equilibrium situations in which electrons and ions exhibit different temperatures. Using machine learning to construct interatomic potentials when materials are not in their electronic ground state is challenging. Here, a neural network interatomic potential is constructed for laser-excited silicon, which extends first-principles accuracy to ultra-large length and time scales.

Automated summary

Data-driven interatomic potentials based on machine-learning approaches are widely used for large-scale simulations of materials. However, these potentials are limited when it comes to describing situations with excited electrons. In this study, a neural network potential with explicit dependence on electronic temperature is proposed. The potential accurately reproduces important physical properties and allows for simulations of laser-induced surface modifications on a thin film at ultra-large length and time scales.