I am about 3 weeks into run time on the program and was trying to figure out if this is about standard or if I need to upgrade to a better computer for such calculations.
@Linda: As others already pointed out, the time you observe is very unusual for a standard single point computation. By the way: what do you mean by convergence? The HF part of your calculation (MP2 is non-iterative and therefore does not have to converge)?
If you do a geometry optimization (which of course has to converge), the run time could be more likely, depending on your start geometry etc. MP2 gradients can be quite costly. In that case it might be better to do a pre-optimization with a reasonable DFT functional (including a dispersion correction!).
@Jiří: What do you mean? Using a post-HF method (like MP2) WITHOUT a sufficiently large basis set is useless and a mere waste of time. Maybe Pople basis sets are not the best suited, there are better choices. In that regard I agree. But if you are claiming that LARGE basis sets are not a good idea then I challenge that.
Convergence for what? Single point energy? Geometry optimization? Frequency calculation?
Which program? With TURBOMOLE a single point MP2 energy with 250 bfs would take maybe something like 1 min.
Why do you believe this level is reasonable? I mean large Pople basis set and MP2 is not very good combination in general.
I agree with the previous posters. With ORCA a single point MP2 calculation with 250 basis functions would take some minutes.
I think u should increase the memory (shd be given in the input of gaussian) . by knowing ur primitives u shd calculate necessary memory for that. In gaussian manual u find how to calculate sufficient memory calculation.
I mean u need not go for upgrading a system configuration. but u can try above suggestion. length of time also depends on number of atoms also. if too many atoms it will take lot of time.
@Linda: As others already pointed out, the time you observe is very unusual for a standard single point computation. By the way: what do you mean by convergence? The HF part of your calculation (MP2 is non-iterative and therefore does not have to converge)?
If you do a geometry optimization (which of course has to converge), the run time could be more likely, depending on your start geometry etc. MP2 gradients can be quite costly. In that case it might be better to do a pre-optimization with a reasonable DFT functional (including a dispersion correction!).
@Jiří: What do you mean? Using a post-HF method (like MP2) WITHOUT a sufficiently large basis set is useless and a mere waste of time. Maybe Pople basis sets are not the best suited, there are better choices. In that regard I agree. But if you are claiming that LARGE basis sets are not a good idea then I challenge that.
As for your question, I would like to say that the time needed for the convergence of wavefunction optimization is depends on the properties and scale of your studied system, and mostly different software gives different time length of wavefunction optimization for the same system.
Convergence of the initial wavefunction depends heavily on the starting coordinates, if the system is not in a reasonable conformation (check all bond lengths/angles) it will never converge. Try to use a preminimized input structure at a lower level (force field / DFT).
Good Luck, Marco
My apologies, I should have been clearer. I am doing geometry optimization. When I first began to attempt this (I realize these sort of calculations are not done because of the time demand and size) I was getting very large gradients (in the realm of millions), after using data I had I was able to tweak the system down to a gradient of 15. I did start with molecular mechanics, then increased the level of theory through smaller basis sets and without the MP2. I chose my basis set based on previous papers studying the same molecule but under different conditions.
When I ran everything without the MP2, the results were nearly identical to what we had without a higher level of theory (same above basis). It is the interaction we are interested in, thus applying perturbation. The system itself is large (about 1400 atoms), but (before anyone spits their coffee out), I am able to treat just the 24 I am interested in as QM atoms, and treat the rest as a classical perturbation on the Hamiltonian.
Our previous computers could not even begin this sort of task. I was very excited when I found out the new computer could.
@ Abdul, thank you so much for the paper! I had not found that one yet.
@ Tobias , if that sounds like a reasonable length of time for what we are doing, thank you it is nice to know. I wasn't able to figure out if my computer wasn't good enough (I read this calculation could take up to 20 gigs, don't know how accurate that is) or if what I was asking it to do was just a time consuming matter and I must be patient.
Dear Linda,
if your system interacts via noncovalent interactions, optimzations can take very long and BSSE (basis set superposition error) is huge for small pople basis sets. Preminimization with HF methods will result in totaly wrong starting geometries due to the miscalculation of weak interactions. You may try FF only or DFT + vdW-correction. You may also try local MP2, this is specially designed for larger systems and avoids BSSE.
Especially in interactions, the coupling and embedding sceme is annother source of error. Therefore special algorithms are needed to converge such calculations (Check e.g. CHEMSHELL) This problems cannot be solved by changing the computer, only by designing an appropiate boundary and environment.
hi lin!
if u work with a bio-structure you need not to use the same basis funcs. and level of the theory for the whole of ur system of interest. if so, u should apply ONIOM model. it`s quite good.
Marco,
I have been doing some reading on the errors you suggested. Please correct me if my understanding is wrong, but it appears the errors are generated from overlapping basis sets. Do I not avoid this problem by treating the one part of the system classically (my program essentially treats it as Q/R perturbation on the 1e Hamiltonian)?
I am working on a very small student budget. I have an older versoin of Gauss/Homology/HyperChem, which does not work on the newer computer. Essentially I can afford to "fix" only one thing (the programs are more blazing expensive than the computers themselves which is why I was rather hoping the latter was the bigger issue). But after more reading it appears a more efficient updated program, such as Gauss09, will allow more control of the process (helping with potential errors) and be more time efficient???
Nasser,
The structure of interest is completely treated with the same basis set. I have not chopped the molecule in half and treated it as a hybrid QM/MM. I am treating the entire molecule quantum mechanically and fixing the second portion of the system as essentially a point charge type perturbation. I want to see the effects of the interaction on the geometry optimization. Please understand, my level of understanding is that of a student, I mean no offense by this. After reading on BSSE, from Marco's comment, wouldn't treating the entire system with the same basis function introduce error? I thought ONIOM does just what I am trying to do (QM/MM hybrid)? Or am I misunderstanding your comment.
By the way, I thank you all so much for your feedback and comments. It has led to more reading on my part and hopefully a better understanding of the process.
Linda
@Tobias: Well, some cc- basis set might be better choice but you have to take into account the fact that with large basis set (approaching CBS limit) the MP2 overestimates the dispersion a lot, in fact it works "quite well" with small Pople basis set due to cancellation of errors (the overestimation is cancelled by BSSE) - but to be clear this is not what I suggest to use. Good "dispersion" corected DFT (in fact parametrized at CCSD(T)/CBS data) will work better in most standard cases.
Dear Linda,
the BSSE will contaminate you results most for noncovalent interactions (H-bonds, vdW) using post-HF QM Methods. Including point charges for distant atoms will not affect this, but will complicate the convergence and reliability of the calculation even more. For compensation of BSSE at low basis set level you will need to calculate the system divided into pseudo monomers plus pseudo-orbitals. (see Jirji's commmment for a proper CBS-limit extrapolation, for which you need CCSD(T) to get reiliable results)
If you do not have the right software and Computer cluster to do this properly, you are in deep trouble anyway, since unreliable MP2 calculations with small basis sets will cause more trouble than they solve. You should discuss this very carefully with your supervisior and the local Theorethical chemistry experts who should have access to a computer cluster and the appropiate software.
All the best,
Marco
@Jiri
Unfortunately BSSE overestimates the interaction and MP2 is overestimating the dispersion, so there is NO compensation at low basis set level. It only get worse.
This compensation works only for B3LYP or HF/631G* for simple H-bonds, due to miscalculation of electrostatics and miscalculation of dispersion and BSSE.
"The right result for the wrong reason"
@Jiří: I completely agree with Marco. At the complete basis set limit there is no BSSE, so this problem is worst for small basis sets. And yes, MP2 does overestimate dispersion in some cases, i.e. when aromatic systems are involved. On the other hand it works almost perfect for alkane chains.
@Linda: Your ideas about BSSE are in principal correct. As long as you do not have two QM fragments (with some overlap), BSSE is less of a problem. But be aware that there exists also the intra-molecular BSSE. It's smaller in normal systems - but there are known exceptions. And on the bad side: There is no way to estimate its effect not to correct it afterwards. On more reason why you need sufficient large basis sets for MP2.
Regarding the choice of software: For really efficient computations (DFT and MP2) I suggest Turbomole, but it's not for free (though not too expensive for academia). Otherwise try ORCA: it's free, contains most of the modern methods, and is easy to use. It also comes with a quite helpful manual with a lot of background information. It has been applied in various QM/MM studies. And finally, there is a recent paper by Neese and co-worker (the developer team) about efficient optimizations with RIJCOSX-MP2 (http://pubs.acs.org/doi/abs/10.1021/ct100199k).
@Marco and Tobias: The point is not to correct for the BSSE, then you have the "compensation". It kind of works for example for 6-31G* or its scaled version 6-31G**(0.25,0.15). But I agree that it would be the "right" answer from wrong reason. No doubt that CCSD(T)/CBS would be much much more reliable. Or from practical reasons, scaled MP3 (MP2.5 or MP2.X)/CBS - you can do the optimization at this level, btw. Anyway, if interested, look for papers published by Hobza (and myself).
Thank you all again!
I have abandoned my easy to use GUI programs in favor of using G03 alone (although I have added Gauss View to my Christmas wish list), and in the meanwhile found papers working with very similar systems; doing what I am trying to do; and they also addressed the BSSE issues!
No doubt, I have more reading to do, but I feel I am off to a nice start. @Jiri, I will search for papers under yours and Hobza's names. I have been reading eveyrthing suggested which has helped in finding other related papers as well.
I have opted to run ONIOM on Gauss with different basis sets than I had originally begun with (MPW1B95/6-31++G**:MPW1B95/3-21G*), but I want to do more reading before asking any more questions which pop up.
Again...I cannot thank you all enough for your input!!
Linda
@Jiri and Tobias,
I have now read some of your papers (along with Hobza and several others from posters here). Can either of you (or anyone for that matter) please tell me, if a functional and basis set is said to give a favorable geometry while underestimating binding energy, is this a red flag for BSSE? If it is, I am looking at how to load the B2-PLYP with Gauss 03 and run it with the TZVP basis set. Does this sound more logical?
Thank you sir! I am up and running within the limits of G03W and loading GROMACS to go with ORCA. (hit a bit of a problem with the fftw) I am very pleased that I asked the quesiton. The feedback has been tremendously helpful!
I strongly recommend some CUDA based software such as TeraChem. You will see a huge performance increases, up to 100x !! Generally you will speak for hours not for weeks:)
Filip,
Thank you for your reply. We were able to get a grant to upgrade our software (GaussView//Gauss 09) and computer. I have had the opportunity to work with current basis sets and functionals with trial software (which takes advantage of multiprocessing). We were pleased with the quality of the results and thrilled with the time saving!
All of the suggestions made here sure helped get us on the right track for finding inavluable research papers. I feel like a kid at Christmas waiting for our new lab to be up and running!!
Thank you all again.....
- Badges
- Science topic
- Similar topics
- Chemistry
- Theoretical Chemistry