ml-property-predict-scd

What problem does it solve?

Enables training or transfer of property-prediction models from SelfConditionedDenoisingAtoms (SCD) foundation checkpoints, reducing the effort to build ML pipelines for atomistic materials and molecular property datasets.

Core Features & Use Cases

• Frozen SCD encoder embeddings: reuse pretrained SCD to generate mol_emb graph-level features (and optionally atom_embs) for downstream ML.
• Lightweight head adaptation: train scalar_head, atom_emb_mlp, or mol_emb_mlp while keeping the SCD backbone frozen for faster iteration.
• Full-model finetuning or pretraining: run upstream train.py for full finetuning or pretraining from scratch on new datasets.
• Dataset onboarding guidance: provides a dataset contract for returning torch_geometric.data.Data with required fields for scalar/energy-force/periodic tasks.

Quick Start

Use the SCD frozen-backbone embedder to produce mol_emb features for a new molecular dataset task.