Molecular LPGM

Reference implementation: https://github.com/MurrellGroup/molLPGM

Fragment latent process generator matching and atom-wise Brownian flow matching for molecular generation. Both models receive full fragment membership conditioning and use the same backbone and training schedule.

Dataset	Width	Layers	Fragment LPGM	Brownian FM
GEOM-Drugs	768	16	geom/fragment.pt	geom/bm.pt
QM9	512	12	qm9/fragment.pt	qm9/bm.pt

All checkpoints contain final 800,000-update raw weights, EMA weights (decay 0.9999), architecture, process parameters and loss weights. Optimizer state is omitted. Use the reference implementation's checkpoint loader with use_ema=True, or pass --ema to its sampling scripts, to reproduce EMA sampling.

Fragment LPGM uses fragment centroid, rotation and internal-coordinate dynamics. BM uses an atom-wise Brownian bridge with source standard deviation 3 Å and a Kumaraswamy clock with a=2.3, b=3.1 and beta=14.8.

Download URLs, SHA-256 hashes, file sizes and parameter counts are in checkpoints.json.

QM9 evaluation

Final EMA samples, using 2,000 square-root-spaced sampling steps and the same 1,024 held-out test cases for both models. All requested samples are included, before viewer filtering.

Model	Valid	Connected
Fragment LPGM	1,010 / 1,024 (98.6%)	1,016 / 1,024 (99.2%)
Brownian FM	1,012 / 1,024 (98.8%)	1,016 / 1,024 (99.2%)

Validity uses OpenBabel bond perception followed by RDKit sanitization and a canonical SMILES roundtrip. Connectivity is assessed independently and requires one reconstructed component. See the reference implementation's model and training details.