My Community

Dear everyone,
I want to use machine learning to train a potential function for SiC at 2000K. First, I turned off machine learning for AIMD testing, but the temperature fluctuations in my OSZICAR file are very severe. I would like to know how to solve this problem. The following is the setup of my INCAR file.
SYSTEM = Si C
ISMEAR = 0
SIGMA = 0.1
ISYM = 0
NELM = 100
EDIFF = 1E-06
LWAVE = .FALSE.
LCHARG = .FALSE.
EDIFFG = -0.01
#MD
IBRION = 0
MDALGO = 3
ISIF = 3
PSTRESS = 0
SMASS = 3
APACO = 16
TEBEG = 2000
TEEND = 2000
NSW = 200
POTIM = 0.5
RANDOM_SEED = 88951986 0 0
#Machine-learned
#ML_LMLFF = .TRUE.
#ML_MODE = train
#ML_ISTART = 0
#Parallelization of ab initio calculations
NCORE = 2
#For long MD-runs use Nwrite = 0 or 1
Nwrite = 0

Are you starting from an equilibrated liquid? If not you need to equilibrate otherwise you will see massive fluctuations.

If you haven't and you start from a solid, you can either wait long with AIMD to equilibrate or use MLFF with on-the-fly learning to speed up.

For the liquid you may also need an ICONST file otherwise the liquid cell may irreversibly deform.

Dear ferenc_karsai,
Thank you very much for your answer! I have been simulating SiC-beta crystals. When conducting AIMD tests, I found that the calculation speed is extremely slow, and the temperature fluctuates drastically in the early stage, which is very time-consuming. Then I tested the case with machine learning enabled. At around 5000 steps, the temperature fluctuates around 2000 K. Can I continue the training in this situation?

You should also visualize your cell/trajectory at that temperature, does it look like a proper melt? Ovito is great for visualizing trajectories. Check also if the volume fluctuates strongly or the cell deforms. As I wrote before an ICONST file is very likely neccessary to keep the melt from collapsing.

You have a step size of 0.5 fs which would result only in 2.5 ps, that might not be long enough.