MolecularDiffusion.modules.models.etflow¶
ET-Flow: equivariant flow matching for molecular conformer generation.
Hassan, Shenoy, Lee, Stark, Thaler, Beaini, NeurIPS 2024, arXiv:2410.22388. Upstream: https://github.com/shenoynikhil/ETFlow (MIT).
Three pieces land here and nothing else:
torchmdThe time-conditioned TorchMD-ET transformer that is the vector field. Module names match the released checkpoints exactly.
flowThe harmonic prior over the bond-graph Laplacian, the Kabsch alignment that makes the regression target rotation-free, the edge-set rebuild and the batchwise L2 objective. Parameter-free.
featuresET-Flow’s 10-column atom featurization and its chiral-centre tensors.
Upstream’s Lightning/optimizer plumbing, dataset code, CLI and evaluation scripts are deliberately NOT ported – the platform engine owns all of that.