How acesuit/mace works
MACE is a PyTorch implementation of equivariant machine-learning interatomic potentials. Its main workflow turns atomistic configurations into graph data, trains a model, and exposes predictions through calculators for simulation or evaluation. The repository also includes foundation-model fine-tuning, polar and magnetic model capabilities, dataset preprocessing, and model conversion utilities. The graph emphasizes documented capabilities and relationships visible in the README or source index; several runtime details are unsampled, so uncertain links are omitted or marked unverified.
Main entry points: run_train.py Training CLI · blocks.py Equivariant blocks · atomic_data.py Atomic graph data · train.py Training engine
Components 23
Atomistic data
- Data preprocessing
mace/cli/preprocess_data.py - Atomic graph data
mace/data/atomic_data.py - Neighbor construction
mace/data/neighborhood.py - Graph batch types
mace/tools/torch_geometric/data.py - HDF5 dataset
mace/data/hdf5_dataset.py - LMDB dataset
mace/data/lmdb_dataset.py
Potential models
- Equivariant blocks
mace/modules/blocks.py - Radial features
mace/modules/radial.py - Specialized models
mace/modules/extensions.py - Training losses
mace/modules/loss.py
Training and analysis
- Training CLI
mace/cli/run_train.py - Training engine
mace/tools/train.py - Fine-tuning selection
mace/cli/fine_tuning_select.py - Configuration evaluation
mace/cli/eval_configs.py
Inference and simulation
- Foundation models
mace/calculators/foundations_models.py - ASE calculator
mace/calculators/mace.py - LAMMPS integration
mace/calculators/lammps_mace.py - Active-learning MD
mace/cli/active_learning_md.py
Model utilities
- Equivariance conversion
mace/cli/convert_e3nn_cueq.py - Backend operations
mace/modules/wrapper_ops.py - Scatter operations
mace/tools/scatter.py
Other components
- MACE user
- Simulation user
Connections 22
- MACE user → Training CLI: starts training
README.md - MACE user → Data preprocessing: prepares data
- Training CLI → Atomic graph data: uses data types
mace/cli/run_train.py - Atomic graph data → Neighbor construction: builds neighbors
mace/data/atomic_data.py - Atomic graph data → Graph batch types: creates graph data
mace/data/atomic_data.py - Training CLI → HDF5 dataset: loads dataset
- Training CLI → LMDB dataset: loads dataset
- Equivariant blocks → Radial features: uses radial features
mace/modules/blocks.py - Specialized models → Equivariant blocks: uses blocks
mace/modules/extensions.py - Training CLI → Training engine: runs training
- Training engine → Training losses: computes loss
- MACE user → Foundation models: uses pretrained models
README.md - Foundation models → ASE calculator: wraps calculator
mace/calculators/foundations_models.py - Simulation user → ASE calculator: requests predictions
- Simulation user → LAMMPS integration: runs simulations
- MACE user → Fine-tuning selection: selects fine-tuning
README.md - MACE user → Configuration evaluation: evaluates configurations
- MACE user → Active-learning MD: runs active MD
- MACE user → Equivariance conversion: converts models
- Equivariance conversion → Backend operations: configures backend
mace/cli/convert_e3nn_cueq.py - Equivariant blocks → Backend operations: uses operations
mace/modules/blocks.py - Equivariant blocks → Scatter operations: aggregates messages
mace/modules/blocks.py