About atoms 3D geometry and distance matrix for QM9 dataset

[xptree]

I am curious that why chainer-chemistry uses rdkit to generate 3D distance matrix instead of using 3D geometry in QM9 dataset directly. I ask this question because the 3D geometry provided by QM9 dataset is believed to be much more accurate than that by rdkit. Do I miss something in the code?

Code details:

• chainer-chemistry skips the 3D geometry when reading QM9 dataset:
  https://github.com/pfnet-research/chainer-chemistry/blob/c05b879241cf62ccd78982145243705b1f81e43b/chainer_chemistry/datasets/qm9.py#L119-L127

• chainer-chemistry calls rdit to get the 3D distance matrix:
  https://github.com/pfnet-research/chainer-chemistry/blob/6f30a6589b591716c0414352d501c9275971b5b5/chainer_chemistry/dataset/preprocessors/schnet_preprocessor.py#L42

[corochann]

Thank you for your interst & question.

I guess DFT need to be executed to calculate precise distance matrix in QM9 dataset, which means that cost for using/calculating precise distance matrix is almost same with calculating labels(energy etc) in QM 9 dataset. Thus we use "rough" distance matrix calculated by rdkit.

And simply, we could not investigate this issue further yet.
#287

[xptree]

@corochann Got it, thanks for your reply.