Spikes in temperature FFT

[gpanchap]

Dear Francois,
I am attaching my input file for Si and the resulting temperature (for the entire run). These are NPT runs. The residual pressure oscillates about zero, as expected for a working barostat. The temperature is set to 1100K. The LOWE time-constant is chosen to be small ~30 au. While the average temp. oscillates close to 1100K taking a FT one sees sharp spikes. As such I am worried if my choice of th_time is perhaps not optimal and the ions are not coupling efficiently with the thermostat. Of course my choice is far from the default value of 5000 au (~120 fs), but I thought this should be fine as the average-T oscillates about the target temperature. Could you please comment on this?

Thanks.

Ganesh

[fgygi]

Hi Ganesh,
I looked at the T(t) data and computed its power spectrum using a maximum entropy method. There is a peak near 1100 cm-1 which would correspond to a vibrational mode at about 550 cm-1 (the frequency of oscillations in the temperature are twice the frequency of the atomic motions). There are also several other peaks at higher frequencies. My interpretation of these peaks would be as follows: the peak at 550 cm-1 may be related to the Raman peak near 520 cm-1 observed in amorphous-Si, possibly shifted in frequency due to pressure. The other peaks are probably spurious and result from the use of the thermostat. Using a high collision frequency (th_time=30) causes many discontinuous changes in the kinetic energy on a short time scale, which likely results in a large spectral content at high frequency.
I would suggest that you try the new Bussi-Donadio-Parrinello thermostat that has been implemented since version 1.52.0. The keyword for that thermostat is BDP. This thermostat yields smooth trajectories, as opposed to the LOEWE and ANDERSEN thermostats. This should result in a better power spectrum of T(t), i.e. without high-frequency noise.
I have the following questions regarding this simulation:
- Did you verify that the system is liquid? Using a Loewe thermostat with very high collision frequency can lead to a large increase in viscosity. This may affect your simulation.

Best,
Francois
-

[gpanchap]

Thanks. I haven't yet analyzed the T=1100K structure. But it may have become amorphous as it is highly supercooled. If so, that would explain the large fluctuations in the energy landscape. I did look at the structure for a 1800K (liquid-T) run. There I had used a th_time~80 and the FFT of the temp. didn't show any spikes. Its g(r) and MSD compared well with a previous VASP run. The average temp. wasn't really 1800K, instead it was close to 1700K, hence I increased my collision frequency for the 1100K run. I will try the new thermostat and see if I can get good results with it at the low-T. Thanks again.

Ganesh

[fgygi]

I have used the following parameters successfully for liquid Si at 1800K. These parameters lead to reasonably good energy conservation in an NVE calculation.

Code: Select all

# liquid Si at 1800K
# load 200-atom sample, ecut=15Ry, nempty=100
load bdp.xml
set wf_dyn PSD
set ecutprec 3
# fermi occupation factor, T=2000K
set fermi_temp 2000
set charge_mix_coeff 0.8
# Bussi-Donadio-Parrinello thermostat
set thermostat BDP
set th_temp 1800
set th_time 4000
set atoms_dyn MD
set dt 40
# remove center of mass motion
reset_vcm
# run 1000 steps, approx 1 ps
run 1000 10 5
save bdp.xml


Reducing the thermostat response time (th_time) can help maintaining the target temperature, which also allows for a less strict convergence of the electronic ground state at each ionic step. However, this should be done with caution since dynamical properties may be affected.