QTAIM-Embed

A GNN package for molecular properties. These models can handle spin and charged species as well as complex atom, bond features. QTAIM features are compatible via a sister package Generator and can add performance and robustness to existing models Currently only structure-to-property models are supported but we are working on structure-to-node level models.

The current implementation supports a good starting set of message-passing functions:

• GCN
• GAT
• GCN+Residual

In addition, several global readout functions are implemented for size-intensive and extensive properties including:

• Mean
• Sum
• WeightedMean
• WeightedSum
• GlobalAttentionPooling
• Set2Set

TODO: some instructional notebooks

Security Notice

LMDB data files and scaler files use Python pickle-based serialization (torch.save/torch.load). Do not load LMDB datasets or scaler files from untrusted sources, as they may contain arbitrary code that executes during deserialization.