Official Resources
- Homepage: https://deeph-pack.readthedocs.io/
- Source Repository: https://github.com/mzjb/DeepH-pack
- Developers: H. Li Group (Tsinghua University)
- License: MIT License
Overview
DeepH is a state-of-the-art framework for Machine Learning Enhanced DFT. It bypasses the computationally expensive self-consistent field (SCF) iterations of traditional DFT by learning a mapping from atomic structures to the DFT Hamiltonian. Utilizing E(3)-equivariant neural networks, it can predict accurate Hamiltonians for large-scale material systems, generalizing from small unit cells to huge supercells with $O(N)$ inference cost.
Scientific domain: Material Science, Machine Learning, Large-scale Electronic Structure.
Target user community: Researchers studying twisted bilayers, large supercells, and complex material interfaces.
Theoretical Methods
- Approach: Supervised learning of the DFT Hamiltonian matrix.
- Architecture: E(3)-equivariant graph neural networks (e.g., DeepH-E3).
- Basis: Compatible with localized basis sets (PAO/LCAO).
- Workflow: Train on small system DFT data -> Predict on large system -> Diagonalize/Calculate properties.
Capabilities
- Prediction: Direct prediction of Hamiltonian and Overlap matrices.
- Scalability: Capable of handling 10,000+ atom systems where conventional DFT is prohibitive.
- Interfaces: Works with OpenMX, ABACUS, FHI-aims, and SIESTA.
- Applications: Moiré superlattices, defects, amorphous systems.
Key Strengths
- Efficiency: Reduces calculation time from hours/days (DFT) to minutes (Inference).
- Generalizability: A model trained on small perturbations can predict large, unrelaxed structures.
- Accuracy: Retains near-DFT methodology accuracy (meV scale) unlike simpler ML potentials which only give energy/forces.
Performance Characteristics
- Speed: Inference is 100-1000x faster than corresponding DFT calculations.
- Scaling: Linear scaling $O(N)$ with system size for inference.
Limitations & Known Constraints
- Transferability: While better than simple ML potentials, preventing "catastrophic forgetting" or ensuring transferability to completely unseen chemical environments remans a challenge.
- Training Cost: Generating the DFT training database is the bottleneck; the ML model can only be as good as the DFT reference (inherits DFT errors).
- Complexity: Predicting a Hamiltonian matrix is dimensionally more complex than predicting a scalar energy, requiring significant GPU memory for training.
Best Practices
- Dataset: Ensure the training set covers the structural deformations expected in the production run.
- Basis: Use a consistent localized basis (e.g., OpenMX PAOs) for both training and inference; you cannot mix basis sets.
Community and Support
- Support: GitHub Issues and documentation on ReadTheDocs.
Comparison with Other Codes
- vs Conventional DFT: DeepH is not a replacement but an accelerator/extender. It needs conventional DFT for training data.
- vs ML Potentials (NequIP/MACE): Potentials predict energy/forces for MD; DeepH predicts the electronic Hamiltonian, effectively giving access to band structures and wavefunctions of large systems.
Verification & Sources
Primary sources:
- Repository: DeepH-pack GitHub
- Literature: Li, H., et al. "Deep-learning density functional theory Hamiltonian for large-scale electronic-structure calculation." Nature Computational Science (2022).
Verification status: ✅ VERIFIED