Is it wise to use ECPs on light atoms for electronic structure calculations?

So, I have a big organic molecule consisting of $\ce{H}$, $\ce{C}$, $\ce{O}$, and $\ce{N}$ atoms and the goal is to find its equilibrium geometry and US/Vis spectrum. The molecule is so big that I was thinking about reducing the complexity of the calculations by using effective core potentials (say, Stuttgart RLC) on light atoms ($\ce{C}$, $\ce{O}$, $\ce{N}$). However, I'm not sure how trustworthy the results would be comparing to using full-electron bases, thus, I'm looking for some reviews if there are any.

• What do you mean by "big?" Is it big enough that linear scaling DFT methods would help? – Geoff Hutchison May 29 '15 at 17:22
• @GeoffHutchison, by "big" I simply mean big enough to make my calculations too slow. TD-DFT, for instance, take few weeks. :D I obviously want calculations to proceed faster but without introducing physically horrible approximations. If it possible, of course. – Wildcat May 29 '15 at 18:45
• @GeoffHutchison, ~300 light atoms. I tried to turn on the use of the fast multipole method in Gaussian right now and I can see a 20% decrease in computational time. But I want more! :) – Wildcat May 29 '15 at 18:50
• you might consider using more nodes. I don't know how parallel TD-DFT is in Gaussian (or other packages) but there is obviously a significant improvement in speed when running parallel calculations on large jobs. – Geoff Hutchison May 29 '15 at 18:53
• @GeoffHutchison, I see almost now speed-up running on 64 cores vs 32 cores in GAMESS and NWChem, ORCA simply fails to run on 64 cores. And Gaussian setup I have access to is limited to 1 node (16 cores). – Wildcat Jun 17 '15 at 8:52

I think you will find that the time savings are minimal. Modern integral screening will eliminate virtually all inter-atom core-core integrals anyway and if you only have a 1s core then the cost of evaluating intra-atom core-valence integrals will be comparable to evaluating the ECP-core interaction.

• Hmmm... I did some benchmarks, of course. Each SCF cycle with Stuttgart RLC ECPs is about 60% faster that with Dunning/Hay "double zeta valence" basis set (DZV) in GAMESS (US). – Wildcat May 21 '15 at 11:28
• Don't know the Stuttgart ECP very well. How similar is it do DH? Same contraction, etc.? Was this with DFT or RHF? – Jan Jensen May 21 '15 at 12:12
• Hybrid DFT, PBE0 to be more exact. – Wildcat May 21 '15 at 13:38
• DZV has (10s, 5p)/[3s, 2p] contraction scheme for C, N, O, while Stuttgart ECP has (4s, 5p)/[2s, 3p]. – Wildcat May 21 '15 at 13:46
• OK, so it's the speed up is probably a combination of 2 things: 1. a slightly smaller basis set and 2. fewer grid points for the DFT. My guess is that the latter is the big factor, which I didn't consider when formulating my answer. As I recall the grid is pretty dense near the nuclei. – Jan Jensen May 21 '15 at 16:53

If you want to make use of symmetry and reasonable parallelization for your TDDFT, you may want to try out ADF.

ADF doesn't use ECPs, but you can freeze the core electrons.

• I think the answer is fine, particularly the frozen core, but please try to avoid "advertising links." – Geoff Hutchison Jun 17 '15 at 12:17
• OK, good to know. It's just that I also saw a link to ORCA. – Fedor Jun 17 '15 at 12:24

The ECP scheme will not bring substantial savings. If the molecule is really large (1000 atoms), you should better think of more efficient approximations, namely RI methods, RI-J, RI-JK or RIJCOSX. You can find their efficient implementation in ORCA package.

In addition, I would suggest using some more recent basis sets, namely the Alrichs def2.

• Thanks, I'm aware of this RI/density fit/greed free DFT, whatever way you call it, but. 1) I'm not familiar with the underlying theory and currently have no time to do so. And I try to avoid using methods if I have no idea how they work. 2) In programs I know how to use RI-DFT is not available for functionals of my choice and I (again) currently not in the position to master a new program. 3) I do see substantial savings and the only thing I want to know is how generally good/bad results of such calculations would be. – Wildcat May 22 '15 at 6:52
• Hello, I can see your point. The speeds up evaluation of integrals and the errors are well documented and known (much better than ECP). The wavefunction parameters are therefore also good approximation and you can use them as excellent guess to run full canonical DFT. For geometry optimizations the accuracy is perfectly sufficient (especially for large molecules). The use of RI is straigtforward, see sites.google.com/site/orcainputlibrary/… – ssavec May 22 '15 at 8:44
• Now at least I know what I gonna do this weekend. :D I'll try to run Orca... – Wildcat May 22 '15 at 14:04
• ORCA in fact is slower than Gaussian/GAMESS. Yes, it benefits from RI trick, but it did not take advantage of molecular symmetry. My molecules are highly symmetric (most of them belongs to $\mathrm{D}_{2}$ point group) and as a result ORCA is slower. – Wildcat May 23 '15 at 10:41
• Besides, ORCA does not have a simple way of keeping symmetry during geometry optimization, so ORCA is by no means a solution. – Wildcat May 23 '15 at 10:43