as a rule of thumb, you need to simulate a time at least 100 longer than the correlation time of your acf, and saving data with at time intervals at least 100 smaller than the same correlation time.

But since you don't know the correlation time beforehand, it means that a few trials could be required.

Dear Luca and Zhaoxi,

I have performed classical NVE simulation at timestep 1fs for 10ns and I want to know vibrational spectra. We have saved the xyz corrdinate at 5 fs and 10fs. Total number of data points I have at 5fs and 10fs are 500000 and 1000000. The result attached below shows a large difference between 5 fs and 10fs recorded data , so now I am confused which time scale 5 fs or 10fs will be more reliable . Kindly give your expert comments .

dear Aashish Bhatt ,

looks like a bug somewhere, but you should check /visualize first the time dependent quantity (dipole for IR spectrum? bond distance of your stretching mode?), then its autocorrelation function, and only after the acf looks converged, you can proceed to its Fourier transform, excluding the long time regions which are normally noisy.

Dear Luca

Please if possible can you share article or document for this problem.so it will be easy to me.

http://manual.gromacs.org/2019/reference-manual/analysis/correlation-function.html

http://www.people.virginia.edu/~lz2n/mse627/notes/Correlations.pdf

https://www.researchgate.net/profile/Mohamed_Elbagerma/post/Can_anyone_explain_the_difference_between_FTIR_spectroscopy_of_sintered_hydroxyapatite_and_TCP/attachment/59d636d9c49f478072ea476a/AS%3A273678711885835%401442261497346/download/33.pdf

https://ocw.mit.edu/courses/chemistry/5-72-statistical-mechanics-spring-2012/lecture-notes/MIT5_72S12_master4.pdf