A deep learning pseudopotential framework for simulating excited state dynamics in semiconductor nanomaterials

Understanding excited-state dynamics of semiconductor nanocrystal (NC) is crucial for novel designs of optoelectronic devices. The semi-empirical pseudopotential method has emerged as an atomistic computational approach delivering ab-initio level accuracy for NC of experimentally relevant sizes. This method simplifies the nuclei and core electrons into a “pseudopotential” experienced by the valence electrons. The current method uses a physics-motivated pseudopotential function form and fits the function parameters to bulk material properties, but this fitting process converges slowly for complex materials.

To address this, we have developed a novel deep-learning framework, replacing the atomistic pseudopotentials with physics-aware deep neural networks. An equivariant descriptor is used to capture the local chemical environment. The deep-learning pseudopotential is trained on ab initio band structures (using DFT+GW) and electron-phonon coupling matrix (using DFPT) to capture both electronic structure and vibronic coupling of the material.

We test our approach with CsPbI₃ and a variety of III-V alloyed NCs. The deep-learning pseudopotentials offer more flexibility to capture the near-core local potentials, outperforming current methods in fitting to band energies and e-ph couplings. It also more accurately predicts key optoelectronic and dynamical properties of NC, including exciton fine structure, linear spectroscopy, non-radiative lifetimes, and electron transfer rates.