# Prerelaxation before AIMD with VASP

I am investigating a bulk structure by ab initio molecular dynamics (AIMD) implemented in VASP. I was wondering, how to choose the initial geometry (= POSCAR file) to run the AIMD simulation. In the literature I saw mainly two different approaches:

1. Using theoretical & unoptimized lattice structures, e.g. from a database like MaterialsProject. This includes lattice constant and a basis (for instance in VASP tutorial https://www.vasp.at/wiki/index.php/Liquid_Si_-_Standard_MD).
2. Using preoptimized lattice structures, i.e. relaxing the ionic positions in before by e.g. conjugated-gradient algorithm (IBRION=2). And then using the CONTCAR as initial geometry for AIMD (I saw this in different papers) .

From my understanding it has more physical meaning, if the structure is relaxed before. Furthermore, lets say I relax a supercell of 2x2x2 unit cells before, I can duplicate the relaxed supercell and use it as an input for AIMD?

Best, Luca

• I really only have experience with classical MD. I was trained to to relax the unitcell first, replicate the structure into a supercell, and then thermalize the supercell. E.g. for NPT, you have to integrate the equations of motion until fluctuations in energy and volume are stable. My experience has shown that skipping the relaxation leads to slower convergence of the volume fluctuations; maybe the statistical sampling is wrong, I am not sure, but I haven't noticed other issues besides longer equilibration. For AIMD, which is expensive, slow equilibration probably matters more. Cheers! Feb 14 at 18:36
• @TySterling thanks for the input. What does stable fluctuation mean ? That the fluctuation value is rather constant, if we compare it for different MD steps ? Feb 16 at 9:35
• Yeah, exactly. For constant P, volume will fluctuate around some mean value. Suppose that the lattice constant in your input structure is too small. During NPT thermalization, the volume of your supercell will increase to correct the specified pressure until volume reaches some "mean" value. It will then fluctuate around the mean. The early stage where the volume is "drifting" rather than fluctuating around the mean is the "transient" stage. The quality of your input structure will significantly affect the transient stage. Some times the transient stage can be very long. Feb 16 at 17:02
• It is worth trying different input structures and plotting the volume over time to really see what I mean. Considering AIMD is very expensive, you could do some testing with classical MD, e.g. LAMMPS, where several thousand atoms can be simulated on a lap top. Then once you have some intuition, you could run your production AIMD calculations. Just a suggestion! Feb 16 at 17:05