Wrong double-bond configuration of generated conformer

[Serious-Huang]

Hi Minkai. Thanks for the nice paper! I am very interested in your work.

Recently, I used GeoDiff (provided trained qm9 model) to generate conformer of a terpenoid compound (Figure 1). And the double-bond configuration of generated conformer is totally wrong, which seems that it's hard to deal with stereochemistry problem.

I'm major in chemistry and unfamiliar with GeoDiff model. Could you explain why the configuration is not the same? And how to deal with it?

I upload the input file (data_test.pkl) and the output file (samples_0.pkl). And I generated the conformer by:

log=log
model=qm9
iter=qm9_default 

python test.py ${log}/${model}/checkpoints/${iter}.pt \
    --save_traj --start_idx 0 --w_global 0.3

file.zip

Any advice would be appreciated. Thanks!