TURBOMOLE Users Forum
Forum General => Miscellaneous => Topic started by: dusan on May 06, 2015, 12:36:35 PM
-
Hi,
I tried to perform a short peptide (~10 amino acids) geometry optimization at the TPSS-D3/def2-SV(P) level of theory using MARIJ.
However, each time I start the optimization (from exactly the same input files) - I get a different geometry. The energy difference between the optimized structures goes even to 0.2 Hartrees.
My guess is that too many degrees of freedom are making a problem for the optimizer. However, is there any simple way to find the closest local minimum?
Cheers,
Dusan
-
Hello,
However, each time I start the optimization (from exactly the same input files) - I get a different geometry.
that should not happen, unless you run into numerical issues:
- do you use internal coordinates?
- is the gradient norm during the optimization very small? Do a 'grep cycle gradient' on the shell or see the gradient output in the result panel in TmoleX. If the gradient is tiny and the energy changes very small, but the max. displacement not converged, the optimizer is running quite blindly on a very shallow PES in a certain direction which could in such cases indeed be driven by numerics.
BUT: I never saw that so far. If you run the same input on the same machine with the same Turbomole version, you should always get identical results. That leads to the next point:
- did you run the job in serial or parallel mode?
- is there a jump in energy during the optimization?
- did you remove the restart files (like hessapprox) after one optimization before you start the next one?
Regards,
Uwe
-
Hi Uwe,
Thanks for the help.
I do use internal coordinates, and I generate them in the define.
The output for one of the optimizations is attached.
I made a plot for 3 different optimization runs from the same starting structure.
In the beginning seems as they all follow the same path, but then around 25th cycle they start differentiating.
(http://s3.postimg.org/k613zyhg3/image.png)
I run everything with TurboMole 6.5 on the cluster. So, the architecture is the same. 12 cores on one node are used with OpenMP.
I noticed this accidentally, so I started testing it. In principle, I generate input files with define and then copy the whole directory several times. Therefore, I don't just restart calculations.
Best,
Dusan
-
You can force the optimizer to run to the closes minimum by setting the maximum step for the geometry update to a small value. But this will usually slow down the optimization.
As Uwe pointed out, the PES is for this problem probably quite shallow. If your amino acids have side groups which can almost freely rotate around single bonds, there will be many local minima seperated by shallow barriers. To get such calculations converged requires a tighter convergence of the density than is default in TURBOMOLE. Try to set scfconv=8 and/or denconv=1.0d-6.
For peptides in gas phase, you should also check the charge states of the end groups. In difference to situation in solution, the zwitterionic structures are not stable in the gas phase and can cause strange problems.