Predictions of Plasmonic Hot Carrier Energies Using Machine Learning

Atomistic simulation of electron dynamics using machine learning provides a pathway to scale computational studies from a few 10s of atoms to device levels with 1000s of atoms, thereby facilitating efficient device design. For example, studies of plasmonic hot carrier-based devices for efficient energy-harvesting has been limited to small scale systems because of the prohibitively expensive quantum-mechanical simulation methods such as nonadiabatic molecular dynamics (NAMD) or real-time time-dependent density functional theory (rt-TDDFT). On the other hand, we have shown that atomistic neural networks (NN) architectures can estimate a time dependent electron density capable of capturing plasmon formation and its subsequent decay into hot carriers, in nanostructures of 500+ atoms, at fractions of the quantum-mechanical simulation time and with minimal quantum-mechanical input data. In this talk, I will present the extension of this work, showing machine-learned hot carriers’ distributions evolving over time as functions of their energy. For this purpose, unlike most NN architectures that focus on scalar predictions, we will introduce a network architecture that predicts orbital occupations transforming with tensor algebra. These ML predictions will enable extraction of useful quantities such as electron-phonon scattering lifetimes that are directly comparable with the experimental measurements. Our goal is to explore the transferability of our workflow in pursuit of a scheme for affordable modeling of hot carrier dynamics in systems with thousands of atoms.