Hello everyone,
I do calculations on compounds with about 25-30 atoms (hydrogen, carbon, oxygen) and about 110 electrons in a water solvent (SMD solvation model). These are optimizations and frequency calculations. I'm searching for the best reference method I could compare results with. These systems are being calculated with B3LYP/6-311**G(d,p), B3LYP/aug-cc-pVDZ and B3LYP/aug-cc-pVTZ.
I was thinking about CIS/6-311++G(d,p) but have found it's not the best pick - it's barely accurate when calculating frequencies. So I came up with an idea to recruit CCSD/cc-pVTZ, though three things are running through my head:
The main problem arises from the technical features of the computer I'm using and time. What's is your opinion? Maybe use something completely different?
Thank you.
Probably it may take a week long or around 10 days...which ultimately depends upon the clock speed of your system...since you are using higher basis functions. There is no other option if you like to go with DFT.
If you don't want to go with DFT you may try with some other force field model or semi-empirical model to calculate your desired property (since you didn't mention what property you are interested in, it is hard to say).
Aadhityan Arivazhagan – Indeed, I missed the idea. I'd like to perform a high-accurate single point energy calculations on reference crystal structure to compare energies from tested functionals with it.
Maciej Spiegel You have try with different functional. Becasue some functional works good on some particular material. Suppose if you want high accuracy, go with MP2 calculation. It may give you better accuracy than the functional you mentioned above.
The functionals in DFT are highly depends upon the atoms and and its structure. Just by atoms we can't acuurately tell wgih functional is effecitive. You may see in the literature of almost same atomic structure and composition and choose according to it. If no literatue available. You have to try on all these...
Yes.If you need a good accuracy, you need to spend more computational time. Try to statrt with more appropriate structure to reduce the computational time.
Dear Maciej,
depending on your molecular system and the property that you want to verify, CCSD(T) with a reasonable basis set is arguably considered as the golden standard in quantum chemistry. While this level of theory is significantly expensive, intensive efforts have been put into developing approaches that reduce the cost of highly correlated QM methods.
In particular, you may want to have a look at the latest implementation of the DLPNO method in Orca 4.2.1 (free; orcaforum.kofo.mpg.de). On page 83 of the manual of Orca 4.2.1, you will find a remarkable example of a single point calculation at the DLPNO-CCSD(T) with a triple-zeta basis set on Crambine, a protein counting more than 650 atoms.
DLPNO-CCSD(T) seems very affordable and should constitute a trustable reference for your case. I tried it for a single point on naphthalene (18 atoms) at the DLPNO-CCSD(T)/def2-TZVP level (def2-TZVP/C auxiliary basis) using all 4 cores of an i7-7700 CPU and about 4GB of RAM. It reached the result in about 15 min. I guess it is very reasonable.
You will find that it is quite consuming in temporary storage (scratch of about 5GB). Since the calculation reads and writes a lot, running it on SSD storage is a good option.
I hope it helps.
Good luck.
Dear Antoine Marion
Thanks for your answe, this sounds really nice.
However, def2-TZVP does not include diffuse function which is important in description of system I'm interested in (number of the hydrogen bonds & pi-electrons spread across the aromatic rings). I do know, there are ma-def2-**P basis sets, or even def2-**PD but these cannot be subbmited with auxiliary basis set /C.
I'll check DLPNO-CCSD(T)/def2-TZVPD.
Kind regards
You can then switch to aug-cc-pVXZ for which there are aug-cc-pVXZ/C auxiliary basis I believe. It might actually be better for this calculation, but that's a little subjective. Alternatively, the "AutoAux" keyword should generate auxiliary basis for any basis set available in Orca. It might take time at the beginning of the calculation, but it should not be much compared to the overall cost.
Good luck.
Antoine Marion , the calculations I specified with ...
DLPNO-CCSD(T) aug-cc-pVTZ/C VeryTightSCF Grid5 FinalGrid6
... have finished. It was really fast – only 40 minutes, even though I increased integration grid and changed convergence threshold.
I hope the results can be considered now as a reference values. Will recalculate them with aug-cc-pVQZ if I have a time.
Thank you for suggesting ORCA.
Dear Maciej Spiegel
I am happy you liked Orca, I also think it is a great software.
However, I think there is a mistake in your input. You did not specify the main basis set, but only the auxiliary basis.
You should specify both:
DLPNO-CCSD(T) aug-cc-pVTZ aug-cc-pVTZ/C VeryTightSCF
Otherwise, the default basis set will be used (i.e., def2-SVP; I just tried it to confirm).
VeryTightSCF is a good idea, you should keep it.
You can further play with the accuracy of the PNO part with keywords LoosePNO, NormalPNO and TightPNO.
Also, the grid parameters should only affect DFT calculations. I think it is ignored for HF and post-HF, and you can remove them.
Finally, this website might be a good starting point to justify your choice and decide if you can trust or not the accuracy of the approximation for your particular case.
https://www.faccts.de/dlpno
Good luck.
Dear Antoine Marion,
Thank you for corrections! Indeed, grid affects only DFT... TightPNO may be relevant for my studies so I will check it.
Seems like ORCA become a friend of mine, though I was focusing on GAMESS as G16 free alternative. Interestingly I didn't find how ORCA estimates required memory and CPUs. My laptop has 4 cores and 16GB RAM – would love to set e.g. 2 cores with 12GB to obtain results faster; especially while I'm not using it.
Dear Maciej Spiegel
The total RAM allocated is controlled with e.g.,
%maxcore 2048
This defines the RAM per CPU core (in MB).
So if you combine that with !PAL2, that will make 2GB/core x 2cores = 4GB.
Just make sure it does not end up with a total greater than your physical memory.
- Badges
- Science topic