Machine learning empirical pseudopotentials for total energy and electronic energy bands

We present a new empirical computational method using neural network as accurate as ab initio methods based on density functional theory (DFT). Our method combines the merits of empirical pseudopotential method (EPM) with machine learning (ML) approaches by introducing a new hybrid descriptor reflecting local symmetries of materials reliably. We demonstrate the versatility and efficacy of our approach by computing energy bands of polymorphs of GeSe and GeSbTe compounds as well as their amorphous structures. Furthermore, we demonstrate that our ML-EPM method can reliably compute total energies and equation of states of materials without learning them separately. Since transferable empirical pseudopotentials in our study can replace all local Hartree, atomic, and exchange-correlation potentials in Kohn-Sham Hamiltonians without cumbersome self-consistency and with significantly reduced plane wave basis sets, results can be applied to all post-processing computational tools within existing first-principles calculations packages with drastically reduced resources compared with conventional DFT.