My Community

I have tried to explore the ferroelectric transition in KNbO3 (NpT with Langevin Thermostat) using ML force field. The 3*3*3 perviskite supercell with 2*2*2 KPOINTS was used to conduct the simulations. I have tried serveral times, the ML force field always generate low ERR and speed up the simulations well. However, I found the phase transition was incorrect severely. The ground state of KNbO3 shoud be a rhomhohedral phase, and the simulations transfered to orthorhombic phase even at a extreme low temeprature (10 K or 50 K) and the rhombohedral phase never existed. The correct phase transition temeprature (rhombohdral to orthorhomic) should be 273 K in experimental. I konw that the MD simulations usually underestimate the ferroelectric transition tempertuare. How can I get a good ferroelectric phase transition order?

INCAR is attached:
# ISTART = 0
# ICHARG = 2
PREC = Normal
EDIFF = 1.0e-04
NCORE = 8
ENCUT = 520
NSW = 40000
ISYM = 0
POTIM = 2
NBLOCK = 1
KBLOCK = 50
IBRION = 0
ALGO = Fast
ISIF = 3
NELMIN = 6
ISPIN = 1
NELM = 60
LASPH =.TRUE.
METAGGA = SCAN
#AIMD
MDALGO = 3
LANGEVIN_GAMMA = 10.0 10.0 10.0
LANGEVIN_GAMMA_L= 100
#PMASS = 30000
TEBEG = 10
TEEND = 10
#RANDOM_SEED = 343529 0 0
#Machine learning paramters
ML_LMLFF = .TRUE.
ML_ISTART = 2
ML_CX = -0.2
ML_CTIFOR = 0.002
ML_CSIG = 0.3
#DOS related values:
ISMEAR = -1
SIGMA = 0.05
LREAL = A
APACO = 10
NPACO = 200
LCHARGE = False
LWAVE = False

Thanks in advcance!

Hello!

Machine learning force fields can only be an aid to speed up the MD runs if the underlying ab initio calculation produces the expected phases. First, please try to reproduce both phases without machine learning. If that works, generate a machine learning force field for both phases separately (using ML_ISTART = 1 to combine them). For investigation of phase transitions it is particularly important to obtain low errors of energies, forces and stresses. e.g., ~1 meV/atom for energies and <= 100 meV/Angstrom for forces. You should also use an independent test data set (structures randomly picked from an ab initio MD calculation at the same conditions) to verify these results. Only then, you can go ahead and try to simulate the phase transition in an MD run.

Regarding phase transitions: depending on the system you may directly observe the phase transition in a heating or cooling run. However, there can be strong hysteresis and the actual phase transition temperature cannot be determined accurately in this way. For many systems it may be even impossible to observe the phase change only by heating or cooling. In both cases, more advanced sampling techniques are most likely needed to obtain meaningful results.

All the best,

Andreas Singraber and Ferenc Karsai

andreas.singraber wrote: ↑Wed May 25, 2022 9:03 am Hello!

Machine learning force fields can only be an aid to speed up the MD runs if the underlying ab initio calculation produces the expected phases. First, please try to reproduce both phases without machine learning. If that works, generate a machine learning force field for both phases separately (using ML_ISTART = 1 to combine them). For investigation of phase transitions it is particularly important to obtain low errors of energies, forces and stresses. e.g., ~1 meV/atom for energies and <= 100 meV/Angstrom for forces. You should also use an independent test data set (structures randomly picked from an ab initio MD calculation at the same conditions) to verify these results. Only then, you can go ahead and try to simulate the phase transition in an MD run.

Regarding phase transitions: depending on the system you may directly observe the phase transition in a heating or cooling run. However, there can be strong hysteresis and the actual phase transition temperature cannot be determined accurately in this way. For many systems it may be even impossible to observe the phase change only by heating or cooling. In both cases, more advanced sampling techniques are most likely needed to obtain meaningful results.

All the best,

Andreas Singraber and Ferenc Karsai

Hi, Andreas Singraber and Ferenc Karsai

Thanks for your reply!

Yes, I have tried to conduct the AIMD without ML force field simulated in the total time of 18 ps, and the result still was not good. The wrong ferroelectric phase order is not due to the ML. However, the NVT with a gamma sampling could present a good ferroelectric characteristic. Based on the above result, I think the NpT ensemble is much different with the NVT, maybe some parameters need to be adjusted. Can you give some suggesttions?

Regarding phase transitions: The heating or cooling run indeed cannot give the very accurate phase transition temeprature due to the short simulation time in MD. I will try to the advanced sampling techniques, but first I need to obtain the correct groud state of KNbO3 using NpT ensemble.

Thanks again!

Zhi Tan