Core Workflows

A naive illustration of the DeepH-pack workflow is demonstrated as follows:

Source: arXiv:2601.02938 [cond-mat.mtrl-sci]

DeepH-pack enables the prediction of ab initio Hamiltonians using deep neural networks. The standard workflow consists of three sequential stages:

1. DFT Data Interface Setup: Prepare and format density functional theory (DFT) calculations for neural network training

2. Model Training: Configure, train, and validate deep learning models on electronic structure data

3. Prediction & Interface: Deploy trained models for Hamiltonian predictions and integrate with downstream calculations

Each workflow builds upon the previous stage, ensuring a consistent pipeline from DFT data to deep learning predictions. These guides are intended for both new users establishing their first workflow and experienced developers extending the framework.

• Data Preparation
  • Interfacing to DeepH with DeepH-dock
  • Interfacing to DeepH with legacy interfaces
  • Guidance and notes for dataset generation
• Model Training
  • Overview
  • The raw training data and Graph files
  • Example: DeepH model training on FHI-aims processed H2O-5k molecular dataset
  • Monitoring the training process
• Making Prediction
  • Prepare the Dataset for Inference
  • Inference with DeepH
  • Error Analysis
  • Determine and update the Fermi energy
  • Plotting the band structure
  • DeepH-Accelerated Self-Consistent Calculations