« first day (256 days earlier)      last day (488 days later) » 

14:15
4
A: ORCA crashes because of RI approximation?

wzkchem5What the doc wants to say is that the program may make an automatic choice of the auxiliary basis set (typically def2/J for non-relativistic calculations and SARC/J for relativistic ones, if I'm not mistaken), but in some (not all) cases they may not be suited for your basis set. It is well-estab...

"The auxiliary basis sets generated by AutoAux are essentially always suitable for the orbital basis set that you specified." This is not correct. Automatic auxiliary basis sets are only accurate if the orbital basis is at least of polarized triple-zeta quality. — Susi Lehtola Sep 27 at 13:49
adding autoaux has done the trick, thanks a lot. I have found out though, that I have multiple atoms for which I should/need to use aug-cc-pVDZ-PP. I do a lot of calculations and some of the mols contain Bromine and Iodine. aug-cc-pVDZ is only defined for hydrogen-krypton. Can I just use newGTO Br,Zn,I "aug-cc-pVDZ-PP" end to specify a different basis set for multiple elements at once? — J.Doe Sep 27 at 14:35
@J.Doe The last sentence in your most recent comment looks like a new question that would preferably be asked in a separate post (not in this one) since it doesn't have anything to do with the title of this thread. Also, have you tried to see if your suggested line works with multiple elements at once? — Nike Dattani Sep 27 at 14:51
I have already launched a bash script that is looping over 70 input files and uses 4 out of 6 cores on my system. I didn't want to launch any separate jobs right now in case a lack of resources crashes the current loop. so I thought it would just be a simple, quick question for clarification — J.Doe Sep 27 at 15:17
@J.Doe Since you are only using 4/6 of your cores, please use one of the remaining two to check whether or not newGTO Br,Zn,I "aug-cc-pVDZ-PP" end works with multiple elements at once rather than asking here. If it doesn't work, and you want to know how to modify your input file to make it work, you can ask that in a separate post. I can add more guidance here if necessary. — Nike Dattani Sep 27 at 15:44
@SusiLehtola Thanks for pointing this out. I checked the AutoAux paper (DOI: 10.1021/acs.jctc.6b01041), and it seems that for double zeta basis sets, AutoAux is usually more accurate than hand-tuned aux basis sets for total energies, but less accurate for reaction energies. In the most unfavorable case (MP2 reaction energies of the DC9 benchmark set, Table 5), the average error of AutoAux is 0.16 kcal/mol, while that of the hand-tuned basis sets (def2/JK + def2/C) is 0.06 kcal/mol. I'm not sure if this suffices to show that AutoAux is unsuitable for double zeta basis sets. — wzkchem5 Sep 27 at 15:45
@SusiLehtola By the way, what is your suggestion when one wants to use a double zeta basis set (say aug-cc-pVDZ-PP), but there is no dedicated auxiliary basis set for it? If using AutoAux is not an option due to accuracy concerns, do you recommend turning off RI, or use the dedicated auxiliary basis set of another double zeta basis set, or use AutoAux on a similar triple zeta basis set (e.g. aug-cc-pVTZ-PP)? — wzkchem5 Sep 27 at 15:48
@wzkchem5 yes, generate an auxiliary basis for a triple zeta basis for the same family. A double zeta basis may not have functions of a sufficiently high angular momentum, leading to large RI errors as the one-center approximation is not valid, see doi.org/10.1021/acs.jctc.1c00607Susi Lehtola Sep 28 at 16:47
@SusiLehtola Thanks, that was a nice read! However I think this argument is only valid for specific automatic generation approaches. Suppose we have an algorithm that can, e.g. generate f auxiliary functions even for an sp basis set, instead of generating only up to twice the ang. mom. of the orbital basis set (i.e. up to d in this case). Then we can still get arbitrarily good accuracy if we add more and more aux. basis functions. Of course, one cannot generate the f functions by atom Cholesky decomposition, but one can at least generate them empirically. I wonder why no one has done so. — wzkchem5 Oct 7 at 8:19
@wzkchem5 only valid for all auxiliary generation methods suggested so far... the alternative is to build an auxiliary basis by explicit optimization which is a wholly different thing — Susi Lehtola Oct 8 at 9:08

« first day (256 days earlier)      last day (488 days later) »