comparing tmQMg data and applicability of u-NatQG

[YAY-C]

check the TMC overlap (based of refcodes) and applicability of u-NatQG MPNN+G

• check refcodes and overlap
• check paper for applicability
• what are the pre-requisites

[YAY-C]

@choglass
overlap is about 63% of the tmPHOTO ! ( = 2696 TMCs)

[YAY-C]

@choglass u-NatQG is applicable, but graphs need to be pre-generated (including SPE + NBO analysis)

• graph need to be pre-built using HyGDL
  • needs SPE + frequencies + NBO analysis _to_dict()
  • level of theory used was :
    • PBE-D3BJ/def2-SVP (geometries and freqs)
    • PBE0-D3BJ/def2-TZVP (SPE + NBO)
• once generated graphs can be directly used for training and testing with tmQMg

full list of inputs for graph generation can be found here

In summary to apply the model we need to:

• re-run SPE + NBO from QC DFT software
• create a list of dict() with geometric + electronic data
• generate the graphs using dedicated package (i.e HyGDL)

[YAY-C]

@choglass @briling
unfortunately, as far as I checked the raw results are not published anywhere ... only the codes to reproduce them!

• try restricting the dataset to the overlapping set and retrain

[YAY-C]

@choglass @briling these are the results for the sub-training and comparison to previous results!

if you want to check the raw results and code used, they're here