Training `md17` on a GPU results in `nan`

[hwidong-na]

Hi,

Thanks for sharing a great work. I've tried to obtain results for all tasks in this project, while md17 consistantly gives nan when tarining. Even a small learning rate does the same.

$ uv run main.py task=md17 training.epochs=100 algebra.device=cuda training.lr=0.0000001
>>> Starting Task: MD17 (aspirin)
>>> MD17 aspirin: 50000 samples (40000/5000/5000 train/val/test)
>>> Energy: mean=-406737.25, std=5.99
>>> Force: mean=-0.0097, std=26.5588
Loss: nan | E_MAE: nan | F_MAE: nan | Val_E_MAE: nan | Val_F_MAE: nan | LR: 0.0000:   1%| | 1/100 

BTW, I also found the following paper is relavant to your work as well.

@misc{ruhe2023geometriccliffordalgebranetworks,
      title={Geometric Clifford Algebra Networks}, 
      author={David Ruhe and Jayesh K. Gupta and Steven de Keninck and Max Welling and Johannes Brandstetter},
      year={2023},
      eprint={2302.06594},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2302.06594}, 
}

Best,
Hwidong

[Concode0]

Thank you for the detailed report and for taking the interest to test these modules!

To be honest, the md17, abc, pdbbind, and weatherbench modules are currently in a skeleton/WIP state. I pushed the initial task structures as placeholders about two weeks ago, but I haven't implemented the full data pipelines or conducted stability tests for them yet.

My recent focus has been entirely on refining the core Clifford/Versor engine and validating it against the Feynman benchmarks. As you rightly pointed out, the NaN errors in md17 are almost certainly a scaling issue. Clifford rotors and their exponential maps are sensitive to numerical ranges; the massive energy scales in MD17 (often in the -400,000 range) likely overwhelmed the current normalization logic, leading to numerical explosion.

The paper you recommended, Ruhe et al. (2023) "Geometric Clifford Algebra Networks", looks incredibly relevant. I will definitely study their approach to numerical stability and equivariance to fix the scaling issues in these larger datasets.

I plan to conduct comprehensive testing once I secure more cloud compute resources for these high-dimensional tasks. I’ll provide a full update and a proper patch as soon as the pipelines are stable.

Thanks again for catching this early and for the great resource!
I will be update readme for notice the partial task is not working.