Skip to content

Overview of DeepH@FHI-aims interface

Authors: Yang Li (DeepH part), Zechen Tang (interface part) and Yong Xu (DeepH group leader)

Current status

The Deep-learning electronic Hamiltonian methods (DeepH) package is a deep-learning framework designed for neural-network modeling of DFT electronic Hamiltonians, developed by Prof. Yong Xu and Prof. Wenhui Duan’s group at the Department of Physics, Tsinghua University. After five years of development, we are excited to announce the release of the DeepH@FHI-aims interface, which enables seamless integration with one of the most widely used DFT codes.

A dedicated overview and outlook of this interface can be found in Chapter 8.3 of the FHI-aims Roadmap, which includes several DeepH references and a Zenodo repository containing demonstration interfaces and datasets. Below we briefly summarize the scope of that overview and discuss why interfacing DeepH with FHI-aims represents an important milestone.

Why DeepH? In parallel with many deep-learning methods that focus on modeling potential energy surfaces, DeepH targets the electronic Hamiltonian, the core quantity for DFT electronic structures. From the Hamiltonians predicted by deep learning, a wide range of mean-field-level electronic properties can be derived. Moreover, Hamiltonians naturally contain rich information, aligning well with the data-hungry nature of deep-learning models. Consequently, DeepH can be viewed as a deep-learning counterpart to empirical tight-binding models, yet retaining first-principles accuracy.

Why FHI-aims? A critical factor for DeepH is data quality. FHI-aims is renowned as a highly accurate all-electron DFT code whose carefully designed and extensively validated algorithms ensure exceptional data reliability, providing an ideal foundation for high-accuracy DeepH training. Another notable advantage of FHI-aims is its broad compatibility with beyond-DFT first-principles methods, achieving state-of-the-art performance in many applications. Integrating DeepH with these advanced frameworks promises substantial efficiency gains given their inherent complexity. We therefore regard the DeepH–FHI-aims integration as a significant step forward for both communities.

It is important to emphasize that, despite the dedicated efforts from both teams in preparing the interface and tutorials, the interface remains at preliminary stage in terms of functionality. We warmly welcome discussions, feedback, and ideas regarding its development and improvement.

Structure of the tutorial

The rest part of the tutorials are arranged as follows:

  1. A brief technical guidance for understanding the theoretical backgrounds
  2. A hand-on guide for installing FHI-aims version compatible with the interface, the interface, as well as the DeepH-pack
  3. A hand-on guide for generating DeepH-formatted Hamiltonian and structural files for subsequent training
  4. How to perform DeepH-pack training
  5. Examples for DeepH inference