Choosing smearing method for finetuning pre-trained models

[dominicvarghese]

Hi everyone,

I want to fine tune the foundational model with my VASP-DFT dataset. My fine-tuning dataset consists of PBEsol DFT calculations for
KTaO3 (wide band-gap insulator). For my bulk training data (NPT/NVT runs), I used the Tetrahedron method with Blöchl corrections (ISMEAR = -5) as it is generally preferred for insulators.

I am now generating the isolated atom energies (E0s) required for the fine-tuning process. When I attempt to calculate the isolated Oxygen energy using settings consistent with my bulk calculation (ISMEAR = -5), I obtain a positive total energy, which seems unphysical for a bound atom.

Simulation Details (Oxygen Atom):Box: 15×15×15 Å
K-Points: 4x4x4 (as required by tetrahedron method)

INCAR
ENCUT   = 500       # Matches bulk exactly
PREC    = Accurate
EDIFF   = 1e-8
LREAL   = .FALSE.
ISYM    = 0         # Symmetry off
IBRION  = -1        # Static
NSW     = 0

Tetrahedron (ISMEAR = -5): TOTEN = 0.11612059 eV (Positive/Unphysical?)

With Gaussian smearing,
ISMEAR = 0, SIGMA = 0.20: TOTEN = -0.32510178 eV
ISMEAR = 0, SIGMA = 0.05: TOTEN = -0.09370829 eV

Given that ISMEAR = -5 is ill-defined for isolated atom, is it standard practice to switch to Gaussian smearing (ISMEAR = 0) just for the isolated atoms, even if the bulk data uses Tetrahedron? If switching to Gaussian is the correct approach, how should I select the SIGMA to ensure consistency with the model's expectations?

Any advice on the correct protocol for generating these E0s would be greatly appreciated.

Thanks
Dominic

[kavanase]

Generally one should take the "energy (sigma->0)" value as the energy from a VASP calculation, which is less sensitive to SIGMA.

The choice of SIGMA shouldn't really matter, assuming it's reasonable and appropriate for the system (0.05 or 0.1 eV would be typical choices). You can check the original OMat24 paper / MPRelaxSet settings for specific details.

I also wouldn't do ISMEAR = -5 calculations for isolated atoms. The tetrahedron method for smearing is only more accurate when kpoints are not fully converged, it should make no difference with an isolated atoms calculation.

[dominicvarghese]

Thanks @kavanase for the quick reply. When I look at the energy(sigma->0) value, these are the results:

With Gaussian smearing for the O isolated atom case (with PBE pseudopotential),
ISMEAR = 0, SIGMA = 0.01: TOTEN = -0.02776953 eV, energy(sigma->0) = -0.02005641 eV
ISMEAR = 0, SIGMA = 0.05: TOTEN = -0.08947445 eV, energy(sigma->0) = -0.05090887 eV

In my NPT/NVT runs for KTO, my sample INCAR (for NPT at 300K) was as follows:

ENCUT   = 500
ALGO    = Normal  
NELM    = 60   
EDIFF   = 1e-6   
ISMEAR  = 0      
SIGMA   = 0.05   
PREC    = Normal   
ISYM    = 0     
LREAL   = A
LMAXMIX = 6   

IBRION  = 0
NSW     = 3000
POTIM   = 1.0
TEBEG   = 300
TEEND   = 300   

MDALGO  = 3    
ISIF    = 3    

LANGEVIN_GAMMA   = 10.0 10.0 10.0 
LANGEVIN_GAMMA_L = 10.0
PMASS            = 1000

So considering the settings with which the fine-tuning dataset was created, I should use the energy value at the ISMEAR = 0, SIGMA = 0.05: energy(sigma->0) = -0.05090887 eV so that the model is fine-tuned correctly and do the similar approach for K and Ta atoms?

Is my understanding correct or am I missing something ?

[kavanase]

I'm not sure then. If you look at the isolated atom energies (calculated with PBE in VASP) here or elsewhere, you can see they are 1-2 orders of magnitude different to these values.

[dominicvarghese]

Sorry, there was a mistake from my part of reporting the values with spin polarization. (forgot to clear the previous WAVECAR & CHGCAR)

ISMEAR = 0, SIGMA = 0.20: TOTEN = -1.68193001 eV, energy(sigma->0) = -1.60479885 eV
ISMEAR = 0, SIGMA = 0.05: TOTEN = -1.56623321 eV, energy(sigma->0) = -1.54695042 eV

These values are in the same order of magnitude as here.

Since I did not use the spin polarized calculations in the bulk, should I re-run those including the spin polarization and use this set of energies for the isolated atoms?

[kavanase]

If your system doesn't have spin polarisation (i.e., in this case, isn't magnetic), then no you shouldn't need to run spin polarised calcs for the bulk system.

[dominicvarghese]

So the rule is: keep the bulk non-spin-polarized calculations to match the physical system(non-magnetic), but use spin-polarized calculations with proper MAGMOM for the individual isolated atom energies?

[kavanase]

I wouldn't necessarily call it a rule, but yes in this case that's what I would recommend.

You could also do the bulk calculations with spin polarisation, which should give the exact same results (because your system is non magnetic) but slower (so obviously not recommended).