Hello José

The answer for this question is not so simple. Semi-empirical methods are parametrized, in other words, their results are fitted by a set of parameters, normally in such a way as to produce results that best agree with experimental data. Thus, the best method will depend directly of your compound. For organic compounds, the PM7 method normally works very good for geometry optimization. You can try this one. However, the best answer for you, in my opnion, is to search at literature for benchmarch studies to find the best method that correlates with the class (functionals groups) of your compounds.

I hope it helps.

Hey José!

If you want to use semiempirical methods, the best in benchmarks are usually Grimme's tight-binding approaches like GFN2-xTB, followed by (depending on the very benchmark) reparametrized PM6 (PM6-D3H4).

Please also consider that composite DFT/SVP approaches offer systematically improved accuracy at a still very reasonable cost. So you might want to consider PBEh-3c in this regard.

Specifically for conformer search, we devised the CREST (https://xtb-docs.readthedocs.io/en/latest/crest.html) program, which is free to download. It combines a GFN2-xTB meta-dynamics search and optimization with (ORCA-driven) DFT for final energies.

Cheers

Jan

The question by the OP is somewhat ill-posed. Is the OP looking for the semiempirical technique that gives the highest degree of correlation with DFT, (and if so DFT with what functional)? or is the OP seeking the semiempirical method that is most accurate? If one seeks good correlation with DFT, then DFTB+ seems like an obvious choice. For highest accuracy, then accuracy with respect to what observable? Conformational energies are not physically observable, so there is no absolute standard of accuracy. Conformational energy differences are physically observable. No semiempirical method is going to be generally accurate in the calculation of conformational energy differences across all types of organic molecules. If the scope of interest can be restricted to a class of chemically-similar molecules, the chances of identifying a method that produces consistently good accuracy is greatly increased. Fernando Martins dos Santos Junior and Jan-Michael Mewes have some useful suggestions along these lines.

Karl, thanks for you answer. I am looking for best correlation between DFT and semi-empirical, regardless the accuracy. Obviously, they have to be reasonably good. Jan-Michael Mewes appears to have a good answer.