Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Geometries are not fully optimized #1

Open
diogomart opened this issue Apr 8, 2021 · 4 comments
Open

Geometries are not fully optimized #1

diogomart opened this issue Apr 8, 2021 · 4 comments

Comments

@diogomart
Copy link

Hello,

The maximum force on any atom of each conformer is, on average, 0.03139 Hartree/Bohr, which is about 70-fold larger than Gaussian's convergence criterion at 0.00045 Hartree/Bohr. This was calculated at the B3LYP/6-311G** level of theory for 5748 of the neutral conformers. Among these conformers, the lowest value of the maximum force on any atom of each conformer was 0.012 Hartree/Bohr.

An optimization of omegacsd_CNBPCT/rmsd92, a conformer depicted in Fig. 8 of the paper associated with this repository, using ORCA 4.2.1 and B3LYP-D3BJ def2-SVP (which is the level of theory used for optimization according to the paper), shows that the geometry of that conformer is not fully optimized. The energy decreases by about 7 kcal/mol after three geometry optimization steps, and more steps are needed for a full optimization. The calculated energy of this conformer as available in this repository is -4130.633327915281 Hartree, which is very close (~0.2 kcal/mol) from the energy tabulated in file energies/b3lyp-svp.txt.

@ghutchis
Copy link
Owner

ghutchis commented Apr 8, 2021

The maximum force on any atom of each conformer is, on average, 0.03139 Hartree/Bohr, which is about 70-fold larger than Gaussian's convergence criterion at 0.00045 Hartree/Bohr. This was calculated at the B3LYP/6-311G** level of theory for 5748 of the neutral conformers.

If anything, the paper shows pretty clearly that there is a difference between B3LYP and B3LYP-D3BJ, as well as between basis sets. Moreover, when you switch programs, you switch grids, integral cutoffs and a host of other settings (RI, RICOS, etc.) -- especially with regards to geometry optimization criteria. Almost every QM program has its own criteria for declaring "done" in a geometry optimization, and that can also depend on the program version or optimization method (GDIIS, etc.).

So no, I would not expect geometries optimized w.r.t. Orca v3 B3LYP-D3BJ def2-SVP (Grid4 IIRC) to be optimized w.r.t Gaussian B3LYP/6-311G**

An optimization of omegacsd_CNBPCT/rmsd92, a conformer depicted in Fig. 8 of the paper associated with this repository, using ORCA 4.2.1 and B3LYP-D3BJ def2-SVP (which is the level of theory used for optimization according to the paper), shows that the geometry of that conformer is not fully optimized.

The geometries are from https://doi.org/10.1002/qua.25512 - which used Orca 3. While ~7 kcal/mol seems quite high, I also wouldn't be surprised if some of the geometry optimizations are not completely converged - there were over 7,000 optimizations. We did check for completion and re-submit as needed, but when performing batch calculations, there are undoubtedly issues.

In any case, the main point of the paper (and the repository) is to provide a set of single-point conformer geometries and energies on meaningful molecules.

In this particular set, we're certainly not going to re-optimize anything, since all corresponding single-point energies for that geometry would need to be re-computed.

We can certainly investigate tighter geometry optimization criteria for the next benchmark.

@diogomart
Copy link
Author

I'll re-optimize with Gaussian and maybe run the DLPNO single point. I'll make that data available. This is a really nice dataset and the fully optimized geometries may be useful to others.

@ghutchis
Copy link
Owner

ghutchis commented Apr 8, 2021

If you want to re-optimize, I'd highly suggest using B3LYP-D3BJ or -D4. There's minimal overhead and the geometries are significantly better. I'd also definitely go with def2-SVP or def2-TZVP family instead of the Pople basis sets.

@diogomart
Copy link
Author

I'll use D3BJ. I never used the def2- basis sets but will definitely consider them. Thanks!

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

2 participants