# Frequency run does not converge in Gaussian16

I have been optimizing some particularly tricky structures with Gaussian. The optimizations end well and they don't have negative frequencies. However, when checking the convergence of the frequency calculation, I have a disagreement:

-optimization run
Maximum Force            0.000029     0.000450     YES
RMS     Force            0.000006     0.000300     YES
Maximum Displacement     0.001422     0.001800     YES
RMS     Displacement     0.000253     0.001200     YES

-frequency run
Maximum Force            0.000030     0.000450     YES
RMS     Force            0.000006     0.000300     YES
Maximum Displacement     0.027710     0.001800     NO
RMS     Displacement     0.003793     0.001200     NO


I have read on the Gaussian website that one should run again the calculation to find the 'true' minimum. However, even if I run the calculation with ReadFC I can only achieve:

         Item               Value     Threshold  Converged?
Maximum Force            0.000005     0.000450     YES
RMS     Force            0.000001     0.000300     YES
Maximum Displacement     0.002793     0.001800     NO
RMS     Displacement     0.000361     0.001200     YES


Since I won't be doing IR spectra calculations but rather TD-DFT calculation, should I be so worried about this or it is ok to trust the less accurate Hessian of the optimization run?

• Changing geometry along the low-frequency modes almost never changes wavefunction so drastically that it would somehow affect TD DFT computations, so you surely do not need to worry about convergence in this case. Also, considering you do not have imaginary frequencies, maximum displacement is just slightly above the threshold, so it's also OK — you can just mention in SI, that one of your system did not meat the standard convergence criteria but was pretty close. Jul 12, 2023 at 15:22
• However, if you have a time and computer resources, it's a good idea to fix this. The easiest way is to run optimization with Opt=CalcAll option, which computes frequencies for each optimization step and in 90% of complex cases gives a nice output, although it's time demanding. It's also a good idea to slightly randomly shuffle atomic coordinates (+/- 0.02 angstroems) before the re-optimization — this could sometimes help to escape those "unsuccessful" local minima. Jul 12, 2023 at 15:32
• Thanks Ivan. Actually, I tried to do a check by lifting a little the optimization thresholds (from the opt=tight to the default opt), I could obtain a 'Stationary point' check at both the optimization and frequency step. I guess that would be alright? Jul 12, 2023 at 19:40
• Actually, that's a much better idea than to start with CalcAll option :-) Jul 13, 2023 at 19:56
• Thanks for the discussion! If you add a summary of this as an answer I will mark it as a reply. Jul 14, 2023 at 6:54