DeepH-hybrid

This code provides an implementation of the DeepH-hybrid method described in the paper A deep equivariant neural network approach for efficient hybrid density functional calculations (arXiv:2302.08221).

DeepH-hybrid is built upon DeepH series of Deep-learning Hamiltonians. The case studies in DeepH-hybrid is carried out with DeepH-E3, but may also be integrated with DeepH-pack. The current code contains only the newly included code of DeepH-hybrid, including a brief demo for these parts.

Installation

The newly added code is executable under the same python environment of DeepH-E3. In case of running this code under a different environment, a Python > 3.9 is required, with following packages installed:

• numpy
• h5py

Usage

The usage of DeepH-hybrid is in full analogy with DeepH-E3. However, here's a few things to note:

Preparation of DFT-hybrid Data

Preparation of DFT-hybrid data for training DeepH-hybrid involves using the ABACUS package. The ABACUS interface for DeepH, initially developed for DeepH-hybrid, is already open-sourced in the DeepH-pack.

For usage of ABACUS, please refer to the manual of the ABACUS package. For interfacing ABACUS with DeepH-hybrid, the out_mat_hs2 option of the ABACUS must be set to 1.

Modifying the $R_{\text{C}}^{\text{hyb}}$

The modification of the cutoff radius, $R_{\text{C}}^{\text{hyb}}$, is done before the training process. A config file must be specified for determining $R_{\text{C}}^{\text{hyb}}$, with an example in example/input/rc_config.json. The config is stored in a json file. Each key and value of the json files specifies the atomic indices (1 for H, 2 for He, etc.) and the cutoff radius for this element (length unit in Bohr). The demo takes the value of $R_{\text{C}}^{\text{hyb}}=2R_{\text C}$, which is used in cased studies of the main text.

A simple example use is provided in the example directory. Just execute bash run.sh at that directory, and the example/input directory will be modified into files in the example/output directory.

The current version of DeepH-E3 supports DeepH-hybrid usage after this modification is accomplished between deeph-preprocess and deeph-train.