NEB calculations using MLFF

[melissa_sanseverino]

Hello! I am attempting to use a ML_FF file to perform NEB calculations. I am running into some issues because I am trying to make two pictures (I have 4 subdirectories) but the calculations are only being run for the first (01) image. I am wondering if something is wrong with my INCAR or if MLFF is not compatable witht this type of NEB calculation. I have attached the results form my run minus the ML_FF file that was in the directory and in each subdirectory. I have also removed the large output files (WAVECAR, CHGCAR) from the 01 subdirectory.
Thank you so much!

[alex]

Hello Melissa,

IBRION = 5 switches on numerical frequency analysis. So you'll not coming to the NEB part at all.

Cheers,

alex

[melissa_sanseverino]

Alex,
Thank you for the advice! I just changed the INCAR to IBRION =3 but the NEB calculation is still not running for in the subdirectory 02.

[marie-therese.huebsch]

Hello,

In principle, NEB works with MLFF. From the INCAR file you have provided, it seems you are using Transition State Tools for VASP which are provided by Henkelman group (and not by VASP).

We can offer help for the VASP internal implementation of nudge elestic bands, see the Wiki documentation or the tutorial. However for support for TST VASP please contact the Henkelman group.

Best regards,
Marie-Therese

[melissa_sanseverino]

Marie-Therese,
Thank you so much for your response! I am currently trying to use the VASP internal implementation of NEB. I have been using the tutorial to understand the NEB calculations but there is no example of an INCAR that uses both NEB and MLFF. Is it possible to use both at once?(compute NEB calculations using a previously developed ML_FF file)
Thank you again!
Best,
Melissa

[marie-therese.huebsch]

Hi,

I have used the Si vacancy hopping as a test case and followed the following workflow:

• I trained a force field on image 02 using the following INCAR

  Code: Select all

  ISYM   = 0        ! no symmetry imposed
  
  ! ab initio
  PREC   = Normal
  IVDW   = 10
  
  ISMEAR = -1      ! Fermi smearing
  SIGMA  = 0.0258  ! smearing in eV
  
  ENCUT  = 300 
  EDIFF  = 1e-6
  
  LWAVE  = F 
  LCHARG = F 
  
  LREAL  = F 
  
  ! MD
  IBRION = 0        ! MD (treat ionic degrees of freedom)
  NSW    = 1000    ! no of ionic steps
  POTIM  = 2.0      ! MD time step in fs
  
  MDALGO = 3                ! Langevin thermostat
  LANGEVIN_GAMMA   = 1      ! friction
  LANGEVIN_GAMMA_L = 10     ! lattice friction
  PMASS  = 10               ! lattice mass
  TEBEG  = 400              ! temperature
  
  ISIF   = 3        ! update positions, cell shape and volume
  
  ! machine learning
  ML_LMLFF  = T 
  ML_MODE   = train
  
  RANDOM_SEED =         688344966                0                0   
  
  

• Then, I refit with the above INCAR file using

  Code: Select all

  ML_MODE   = refit

• Finally, I copy the ML_FF to each subdirectory of the NEB calculation and use

  Code: Select all

  System = Si
  EDIFFG = -0.04
  
  ! DOS related values
  ISMEAR = 0                   # Gaussian smearing
  SIGMA = 0.05                 # Smearing width
  
  ! Ionic relaxation
  NSW = 100                    # Max number of ionic steps
  IBRION = 1                   # Quasi-Newton algorithm
  POTIM = 0.8                  # Step width
  ISIF = 2                     # Cell shape and volume fixed, ions free
  
  ! NEB 
  IMAGES = 4                   # 2 intermediate geometries for the NEB 
  SPRING = -5                  # Spring constant
  
  
  ! machine learning
  ML_LMLFF  = T 
  ML_MODE   = run 
  

The calculation goes through without errors and even produces reasonable results, although I clearly did not train the force field very well. I also want to point out that force fields are good at interpolating but not extrapolating. So you must train in all phases that you want to describe. If you are just interested in getting the reaction path, it may be more efficient to do an ab initio NEB and not ML_FF.

I used vasp_std version 6.5.1 on cpus. Can you please try to reproduce this?

Best regards,
Marie-Therese