How lipengfei1409/deepmd-kit works
DeePMD-kit builds machine-learning interatomic potentials from reference data, supports training from scratch or fine-tuning pretrained checkpoints, then validates and exports models for simulation. Its main workflow spans data preparation, backend-specific training, model evaluation and optimization, and deployment through Python/native APIs or molecular-dynamics engines. Model architectures support energy and force prediction as well as properties such as dipoles, spin, and density of states. The graph reflects README-documented capabilities and sampled entrypoints; unsampled internals and integration wiring are kept general.
Main entry points: deepmd/main.py CLI dispatch · training.py Optimization · deep_eval.py Model inference · entrypoints/main.py PyTorch training
Components 22
Workflow entrypoints
- CLI dispatch
deepmd/main.py - Pretrained models
deepmd/pretrained - DPA-ADAPT
dpa_adapt/main.py
Training and models
- Training data
deepmd/pt/utils/dataset.py - PyTorch training
deepmd/pt/entrypoints/main.py - JAX training
deepmd/jax/entrypoints/main.py - Paddle training
deepmd/pd/entrypoints/main.py - Experimental training
deepmd/pt_expt/entrypoints/main.py - Model architecture
deepmd/pt/model/model/dp_model.py - Atomic descriptors
deepmd/pt/model/descriptor - Prediction heads
deepmd/pt/model/task - Training losses
deepmd/pt/loss - Optimization
deepmd/pt/train/training.py
Inference and validation
- Model inference
deepmd/infer/deep_eval.py - Model testing
deepmd/entrypoints/main.py - Backend conversion
deepmd/entrypoints/convert_backend.py - Model compression
deepmd/pt/entrypoints/compress.py - Freeze and export
deepmd/pt/entrypoints/main.py
Deployment integrations
- Python and native APIs
deepmd/calculator.py - MD engines
- ASE integration
deepmd/pt/utils/ase_calc.py
Other components
- Model user
Connections 31
- Model user → CLI dispatch: runs commands
- CLI dispatch → Pretrained models: resolves checkpoints
- CLI dispatch → PyTorch training: dispatches training
- CLI dispatch → JAX training: dispatches training
- CLI dispatch → Paddle training: dispatches training
- CLI dispatch → Experimental training: dispatches training
- CLI dispatch → Model testing: dispatches tests
- CLI dispatch → Backend conversion: dispatches conversion
- CLI dispatch → Model compression: dispatches compression
- CLI dispatch → Freeze and export: dispatches export
- Pretrained models → PyTorch training: initializes fine-tuning
- Model user → DPA-ADAPT: starts adaptation
- DPA-ADAPT → PyTorch training: configures training
- Training data → PyTorch training: supplies batches
- PyTorch training → Optimization: runs optimization
- JAX training → Optimization: runs optimization
- Paddle training → Optimization: runs optimization
- Experimental training → Optimization: runs optimization
- Optimization → Model architecture: updates parameters
- Model architecture → Atomic descriptors: computes representations
- Model architecture → Prediction heads: predicts targets
- Prediction heads → Training losses: produces predictions
- Optimization → Training losses: optimizes against
- Model user → Model inference: loads model
- Model inference → Model testing: evaluates predictions
- Model inference → Python and native APIs: serves predictions
- Backend conversion → Model inference: converts model
- Model compression → Model inference: optimizes model
- Freeze and export → Model inference: exports model
- Python and native APIs → MD engines: serves simulation
- Python and native APIs → ASE integration: supports calculations