Note regarding large errors for all elements (all data)

(Home - PM7 and PM6-D3H4 Accuracy - Manual)

For many elements, there was, and still is, a severe lack of high quality experimental reference data.  A consequence of this is that in order to have a training data set large enough to allow parameter optimization, extra data on geometries and heats of formation were generated using a relatively quick ab-initio method,  the B88-PW91 density functional with a 6-31(G)d basis set.  These data were useful during the development of PM6, however during the development of PM7 strong evidence was found that many of these theoretical reference data, particularly ionized species, were severely in error.  These errors only became obvious when attempts were made to model solids.  For example, when parameters were optimized to reproduce the B88-PW91 reference heat of formation of the hexacyanoferrate ion, [Fe(II)(CN)6]4-, all attempts to model solid potassium hexacyanoferrate gave bad results.  When the ion was removed from the training set much better results were obtained. 

The protocol adopted in generating the tables included reporting all calculated results, unless unambiguous evidence existed that a reference datum was inaccurate, in which case it was deleted from all files. PM6 was parameterized using many ab-initio data, and therefore errors relative to ab-initio reference data are small.  In parameterizing PM7, a lot of ab-initio data was excluded, as they were incompatible with experimental results.  Original 2013 note: Despite the fact that these data were excluded from the parameterization, the protocols adopted in reporting PM7 results required that these data should be included in determining accuracy, and this had the unfortunate effect of making PM7 look worse than PM6.  Hopefully, convincing evidence that the theoretical reference data are indeed inaccurate will soon be developed, at which point the reference data involved can be eliminated. Revised 2020 note: Although PM7 is more accurate than PM6-D3H4 in predicting the properties of some inorganic solids, in general the faults noted earlier were found to be real.