Highlighted by Jan Jensen
Automated and efficient TS searches is difficult and there are only a few benchmark studies out there. But this is the first paper I have come across where they attempt this for organometallics. Given the typical size of organometallic compounds, one needs something faster than DFT for efficiency so semiempirical QM (SQM) methods are the obvious choice as long as these simpler methods can describe the chemistry accurately.
The authors have test MOPAC and xTB interfaces to Zimmerman's growing string method (mGSM) on the 34 unimolecular reactions in the MOBH35 benchmark set. I couldn't find an explanation for the focus on unimolecular reactions but the reason might be that it is easier to geometrically align reactants and products for these reactions.
GFN1-xTB and GFN2-xTB find reaction paths for 31 and 30 reactions, respectively, while the corresponding numbers for PM6-D3H4 and PM7 are 26 and 25, respectively. GFN2-xTB fails to find barriers for 2 reactions with < 1.5 kcal/mol barriers, so if these are discounted then GFN2-xTB performs best.
The TS-guess structures (the highest energy point on the reaction paths) are generally in good agreement with DFT, with heavy atom RMSDs of >0.3Å. It would have been interesting to know how many DFT TS searchers converge starting from the SQM structures. The xTB barrier heights compare reasonably well with DFT, with a MAD of 8-9 kcal/mol.
