How xiaoxun-gong/deeph-e3 works
DeepH-E3 predicts electronic-structure Hamiltonians from material structures. The principal workflow converts DFT data into crystal graphs, trains an E(3)-equivariant model, then evaluates structures to produce predicted Hamiltonians; overlap data can support inference for large structures. Separate capabilities include target analysis, OpenMX preprocessing, and sparse band-structure calculation and visualization. The graph follows the README and source-index references; some CLI-to-kernel and downstream-tool wiring is not shown in the excerpts, so those links are marked unverified or omitted.
Main entry points: kernel.py Training and inference · model.py Equivariant network · data.py Dataset processing · deephe3-preprocess.py Preprocess CLI
Components 27
Model workflow
- Preprocess CLI
deephe3-preprocess.py - Train CLI
deephe3-train.py - Evaluation CLI
deephe3-eval.py - Config parsing
deephe3/parse_configs.py - Training and inference
deephe3/kernel.py - Training utilities
deephe3/utils.py
Equivariant engine
- Equivariant network
deephe3/model.py - Tensor operations
deephe3/e3modules.py - Radial and angular bases
deephe3/from_dimenet/basis_utils.py - Distance cutoff
deephe3/from_nequip/cutoffs.py - Bessel basis
deephe3/from_nequip/radial_basis.py - Tensor-product paths
deephe3/from_nequip/tp_utils.py - Gaussian basis
deephe3/from_schnetpack/acsf.py
Data preparation
- OpenMX parser
deephe3/process_data_tools/process_data.py - Dataset processing
deephe3/data.py - Crystal graph
deephe3/graph.py - Lattice neighbors
deephe3/from_pymatgen/lattice.py
Analysis tools
- Analyze CLI
deephe3-analyze.py - Target analysis
deephe3-target_analysis.py - Result analyzer
deephe3/analyzer.py - Sparse band calculation
inference_tools/sparse_calc.jl - Pardiso band calculation
inference_tools/sparse_calc_pardis.jl - Band visualization
visual_tools/plot_band.py - Band comparison
visual_tools/compare_bands.py
Other components
- DFT data
- Overlap data
- Predicted Hamiltonians
Connections 30
- DFT data → Preprocess CLI: provides input
README.md - Preprocess CLI → OpenMX parser: converts data
- Preprocess CLI → Dataset processing: prepares dataset
- Train CLI → Config parsing: reads config
- Evaluation CLI → Config parsing: reads config
- Preprocess CLI → Config parsing: reads config
- Dataset processing → Crystal graph: builds graphs
deephe3/data.py - Crystal graph → Lattice neighbors: finds neighbors
deephe3/graph.py - Training and inference → Dataset processing: loads data
deephe3/kernel.py - Training and inference → Crystal graph: uses graphs
deephe3/kernel.py - Training and inference → Equivariant network: loads model
deephe3/kernel.py - Training and inference → Config parsing: uses config
deephe3/kernel.py - Training and inference → Tensor operations: registers constructor
deephe3/kernel.py - Dataset processing → Tensor operations: decomposes targets
deephe3/data.py - Equivariant network → Tensor operations: uses operations
deephe3/model.py - Equivariant network → Distance cutoff: uses cutoff
deephe3/model.py - Equivariant network → Bessel basis: uses basis
deephe3/model.py - Equivariant network → Tensor-product paths: checks paths
deephe3/model.py - Equivariant network → Gaussian basis: uses basis
deephe3/model.py - Dataset processing → Training utilities: processes targets
deephe3/data.py - Training and inference → Training utilities: uses utilities
deephe3/kernel.py - Analyze CLI → Result analyzer: runs analysis
- Result analyzer → Training and inference: analyzes results
deephe3/analyzer.py - Target analysis → Training utilities: analyzes targets
- Evaluation CLI → Overlap data: reads for inference
README.md - Evaluation CLI → Predicted Hamiltonians: produces predictions
README.md - Predicted Hamiltonians → Sparse band calculation: supports diagonalization
README.md - Predicted Hamiltonians → Pardiso band calculation: supports diagonalization
- Sparse band calculation → Band visualization: yields band data
- Band visualization → Band comparison: compares bands