Page 1 of 1
Machine Learning Force Field for larger systems
Posted: Wed Nov 12, 2025 5:02 pm
by stefani_setiono
Dear all,
Hope you are well! I have managed to successfully run an on-the-fly simulation for my structure (MOF + water) in its unit cell form. I want to take this force field and run a simulation but with the supercell version and I read on the tutorial site that it is possible to take the ML_FFN and use it to run a simulation of a larger system but I could not get it to work. I am unable to find information about this, which commands to use etc. It would be great if I can get some guidance!
many thanks,
Stefani
Re: Machine Learning Force Field for larger systems
Posted: Thu Nov 13, 2025 7:54 am
by marie-therese.huebsch
Hi Stefani,
There is not really a trick to it. You only need to replace the POSCAR file with the supercell POSCAR. Could you descibe what error or issue appears when you try to run the supercell calculation?
Just as a side remark: I assume the smaller unit cell is also a supercell of the primitive unit cell, right? Because the system you train on should already be of the dimension such that characteristic collective atomic motion is captured.
br, Marie-Therese
Re: Machine Learning Force Field for larger systems
Posted: Thu Nov 13, 2025 11:55 am
by stefani_setiono
Hi Marie-Therese!
Thank you so much for the reply. I tried just changing the POSCAR file with the supercell POSCAR, but is seems like it restarted the whole retraining with on-the-fly method. For context the structure that I trained it on is the 2x1x1 supercell of the unit cell while the structure that I am trying to run now is the 4x3x3 (of the unitcell).
This is my INCAR file:
SYSTEM = GFH2O
ISYM = 0 ! no symmetry imposed
! ab initio
PREC = Normal
IVDW = 12 # or 11 for zero damping, but CP2K by default uses standard damping (likely IVDW=12)
VDW_A1 = 0.4289 # typical value for PBE-D3 (adjust if needed)
VDW_A2 = 4.4407
VDW_S8 = 0.7892
ISMEAR = -1 ! Fermi smearing
SIGMA = 0.0258 ! smearing in eV
ENCUT = 300 !energy cutoff for the plane-wave basis set in eV.
EDIFF = 1e-6 !global break condition for the electronic SC-loop
LWAVE = F
LCHARG = F
LREAL = Auto
! MD
IBRION = 0 ! MD (treat ionic degrees of freedom)
NSW = 100000 ! no of ionic steps
POTIM = 0.5 ! MD time step in fs
MDALGO = 2 ! Nose-Hoover
SMASS = 0.5
!LANGEVIN_GAMMA = 1 ! friction
!LANGEVIN_GAMMA_L = 3 5 10 5 5 ! lattice friction
!PMASS = 10 ! lattice mass
TEBEG = 300 ! temperature
ISIF = 2 ! NVT update positions, cell shape and volume
! machine learning
ML_MODE = run
ML_LMLFF = T
ML_ISTART = 2
ML_WTSIF = 2
ML_MB = 0
RANDOM_SEED = 688344966 0 0
NCORE = 4
Re: Machine Learning Force Field for larger systems
Posted: Fri Nov 14, 2025 7:38 am
by marie-therese.huebsch
Odd, because your ML_MODE=run does not allow for training.
Could you please share a minimal reproducible? In order to look into the issue, we need all input files and some output files like OUTCAR, ML_LOGFILE, stdout.
See posting guidlines.
Re: Machine Learning Force Field for larger systems
Posted: Tue Nov 18, 2025 7:12 am
by stefani_setiono
Dear Marie-Therese,
Attached is the zip file of my output and also my inputs (that I can fit!).
Stefani
Re: Machine Learning Force Field for larger systems
Posted: Mon Dec 01, 2025 8:21 am
by marie-therese.huebsch
Sorry for the long scilence. Unfortunately the forum had a system change and we did not notice that the notification where not send for quite some time. So I totally missed your response. I will look into it this week and get back to you with an update asap.
Re: Machine Learning Force Field for larger systems
Posted: Mon Dec 01, 2025 10:44 am
by marie-therese.huebsch
Hi, I have to come back to you because I am unsure what issue to look for in the example you have sent.
If I understand you correctly, you want to use an existing ML_FF to start a supercell calculation with ML_MODE = run. But the calculation you have sent has ML_MODE = train and ML_ISTART = 0. So it will always start from scratch (as per your input instructions).