How large is the energy drift? Some energy drift is unavoidable, but generally you want to keep it to < 10 meV/atom/ps (1 meV/atom/ps is good).

You're right to investigate the time step, that is the most likely cause of energy drift and because Li is a very light nucleus you will need a small time step. I would have thought that 0.5 fs should be good enough, though of course it depends on the potential energy landscape. Does the energy drift reduce as you make the time step smaller?

You will also need to converge your SCF calculations very well, so that the forces are accurate -- remember that for density mixing the energies are correct to 3rd order in the wavefunction error (using the Harris-Foulkes functional) but the forces are only correct to 1st order.

The only other thing I can think of is whether you're extrapolating the wavefunctions for each new configuration. Does VASP have XL-BOMD? It would be worth trying that if you can, as we find it does make it closer to being simplectic, so aid energy conservation.

All the best,

Phil Hasnip

PS I'm a CASTEP person, not a VASP user, so apologies if any of this is not relevant.

https://arxiv.org/pdf/1006.2710.pdf

Try opening the XDATCAR file in vmd and see what is happening to the structure. It may happen that the bonds are getting disturbed at such high temperatures .

Which potential you are using, if you are using reaxff potential going down to  0.1 fs is needed, plus make sure to increase the temperature gradually with small steps 0.01k per step, check your system for any possible divergence.

fix 2 all temp/berendsen 1 800 0.01

run 80000

unfix 2

maybe smaller timestep, higher ENCUT, or more kpoints needed. Besides, be careful of the SMASS parameters.

How large is the energy drift? Some energy drift is unavoidable, but generally you want to keep it to < 10 meV/atom/ps (1 meV/atom/ps is good).

You're right to investigate the time step, that is the most likely cause of energy drift and because Li is a very light nucleus you will need a small time step. I would have thought that 0.5 fs should be good enough, though of course it depends on the potential energy landscape. Does the energy drift reduce as you make the time step smaller?

You will also need to converge your SCF calculations very well, so that the forces are accurate -- remember that for density mixing the energies are correct to 3rd order in the wavefunction error (using the Harris-Foulkes functional) but the forces are only correct to 1st order.

The only other thing I can think of is whether you're extrapolating the wavefunctions for each new configuration. Does VASP have XL-BOMD? It would be worth trying that if you can, as we find it does make it closer to being simplectic, so aid energy conservation.

All the best,

Phil Hasnip

PS I'm a CASTEP person, not a VASP user, so apologies if any of this is not relevant.

https://arxiv.org/pdf/1006.2710.pdf

Baltej,

If you have tried all of the above suggestions and you still have a large drift in the total energy, could be it that at 800K, the total energy is not conserved? Or perhaps, above the superionic phase transition, the force on the system isn't Central Force any longer?

Dear Gboyega,

The energy should always be conserved in an NVE calculation -- the temperature may change, but the energy should not. Usually what you do is run with an NVT ensemble until the system reaches the target temperature (800 K in this case) and then switch to NVE to study the dynamics with the energy conserved (i.e. without the thermostat interfering with the dynamics). 

Phil

Dear Phil, I am aware of this.... So, the input files are the culprits.

I do not know anything about the VASP package so my comments may be irrelevant. By definition, NVE ensemble must preserve the energy so I guess is that the actual dynamics used in your software package is the culprit. It appears from your posting that the drift gets bigger with time. This is possible if the molecular dynamics you are using only preserves E for short duration but not for long runs. Newton's equations of motion will conserve energy but any approximation such as Runge-Kutta will give rise to violation of energy conservation over long runs.