What it does

MEGNet (MatErials Graph Network) trains graph neural networks to predict properties of molecules and crystals directly from their structures. Feed it a crystal structure or a SMILES string, and it returns formation energies, band gaps, bulk moduli, or a dozen QM9 molecular properties. The project includes pre-trained models serialized as HDF5+JSON files that load with a single from_file call.

The interesting bit

The architecture stacks “MEGNet blocks”—each block prepends two dense layers to a graph network update—allowing the model to tune how far information propagates through the structure. A set2set readout collapses the final graph to a scalar. The authors also found that element embeddings learned from formation energy models transfer well to scarcer properties like elastic constants, which is the kind of practical insight that separates materials ML from generic graph benchmarks.

Key highlights

• Pre-trained models available for QM9 (13 molecular properties) and Materials Project (formation energy, band gap, bulk/shear moduli)
• Published MAEs: 0.028 eV/atom formation energy, 0.33 eV band gap on MP-2018.6.1
• Supports training from pymatgen Structure or Molecule objects with a CrystalGraph converter
• Includes multi-fidelity training examples in a dedicated multifidelity/ folder
• Elemental embeddings exposed for transfer learning to smaller datasets

Caveats

• Deprecated: the authors explicitly state this repo is frozen; active development moved to matgl (PyTorch + DGL)
• TensorFlow/Keras implementation only; the README nudges new users toward the DGL reimplementation instead
• QM9 molecular models are not recommended for molecules outside the training set due to limited coverage
• Some structures with isolated atoms will crash train(); you must filter to valid graphs manually via train_from_graphs

Verdict

Worth a look if you’re reproducing the 2019–2020 papers or need a quick baseline on pymatgen structures without rewriting data pipelines. Skip it if you’re starting fresh—follow the deprecation notice to matgl instead.