TorchRef is a PyTorch-based crystallographic refinement library. It uses reciprocalspaceship (and gemmi) for reflection I/O under the hood.
If you’ve ever wanted to write custom refinement pipelines, or integrate crystallographic refinement into a larger PyTorch model and backpropagate to/from the dataset alongside any other PyTorch-based package, give it a try. Installation details are on GitHub: GitHub - HatPdotS/TorchRef: A library for pytorch based refinement of crystallographic models · GitHub
TorchRef is under active development. If you hit bugs, please open a GitHub issue or post here. If something’s missing or you have a feature idea, let me know or open a pull request. I would be happy about any and all contributions.
FYI: TorchRef’s self-reported R-factors run a bit high (~1–1.5 percentage points) compared to phenix.model_vs_data. The model quality generally should be similar to that of REFMAC5 and phenix.refine.