I am doing a CCSD(T) geometry optimization on a series of molecules. One of the geometry optimizations took about 2 weeks even though an analogous molecule with the same number of basis functions took less then 2 days to converge to a minimum. After 30 iterations, the calculation aborted with the...
+1. Maybe your initial guess was bad? Why don't you optimize the geometry with a cheaper (faster!) method like MP2 or B3LYP, and then use the final optimized geometry as the initial starting point for the CCSD(T) geometry optimization? If you give me the input file I can try to optimize it and see if I get a better result.
@NikeDattani The thing is, my initial guess should be even better than that: I used the optimized geometry from a CCSD calculation as the inital guess for this one. I've included this information in my post.
If you think that fourth step was close to the actual minimum, you could try restarting from there and decreasing the max step size for the optimization, as the default could potentially escape a shallow/narrow minimum. opt=(MaxStep=N) sets the max step to N*.01 Bohr and the default is N=30.
You might also want to include opt=NoTrustUpdate to keep that stepsize throughout the calculation. Have you had any further progress?
I’m voting to close this question because the question has been abandoned. The user has not signed in since July 2nd, and doesn't have any other network accounts, so they won't get pinged to see Tyberius's newest comment. The question can always be re-opened if someone comments saying they have an answer or if the user returns and answers to Tyberius's suggestions.
@Tyberius I tried setting maxstep to 5. It appeared to help at first because I reached a point where 3 out of 4 convergence criteria were satisfied. Even the remaining criterion was very close to the threshold value. The first calculation (with N=30) only produced some points with 2 satisfied convergence criteria. But then the calculation with MaxStep=5 started deviating as well. So I used this point with 3 satisfied convergence criteria as the initial guess for the next calculation with Maxstep=1. A similar thing happened here. I've attached respective output fragments in another comment.
@Tyberius MaxStep=5 'Maximum Force 0.000117 0.000450 YES RMS Force 0.000060 0.000300 YES Maximum Displacement 0.001992 0.001800 NO RMS Displacement 0.001109 0.001200 YES' MaxStep=1 ' Maximum Force 0.000116 0.000450 YES RMS Force 0.000060 0.000300 YES Maximum Displacement 0.001911 0.001800 NO RMS Displacement 0.001007 0.001200 YES'
@NikeDattani My apologies for not being able to respond sooner.
No problem! The question isn't closed yet, and now that you're back and have answered Tyberius's questions, I'm hopeful that the community will get closer to converging on an answer to help you with your problem.
With the maxstep=1, the notrustupdate might help. Depending on the property you are interested in getting from this geometry (for example, reaction energies), it may already be converged enough. Frequencies it would be worth being cautious, but even this close to convergence might be sufficient.
Is CCSD(T) valid for the system? As Nike already asked above, CCSD(T) fails in some cases. If the triples correction is huge, the molecule might not even be bound at the CCSD(T) level!
@QuantumX can you tell us which molecule this is, so that we can tell you whether or not CCSD(T) is legitimate for geometry optimization? CCSD(T) will definitely not work at all geometries. It fails miserably at certain geometries.
