Running MD for metals

[dominic_varghese]

Hi everyone,

I am currently using the on-the-fly Machine Learning Force Field (MLFF) engine (ML_MODE = TRAIN) to generate a training dataset for a metallic system. My goal is to sample the phase space across several different temperatures using NVT and NPT AIMD runs - which require proper choice of ISMEAR and SIGMA values.

In previous literature for this specific reference metal, static structural relaxations were performed using Methfessel-Paxton smearing (ISMEAR = 1 or 2) with SIGMA = 0.010 eV.

However, I am facing a dilemma regarding the best practice for smearing during MLFF dataset generation:

Methfessel-Paxton is generally unsuitable for Molecular Dynamics due to the risk of negative occupancies causing unstable forces and SCF convergence failure.

To ensure the Machine Learning Interatomic Potential maps to a single, mathematically consistent Potential Energy Surface (PES), the electronic smearing (SIGMA) must remain strictly fixed across all training data, regardless of the differing ionic temperatures (e.g., 300 K vs. 1000 K) of the MD runs.

To maintain a consistent PES for the MLFF while ensuring stable MD forces, I am considering using Gaussian smearing (ISMEAR = 0) with a fixed width of SIGMA = 0.020 eV for all calculations—including the initial static geometry relaxations and all subsequent AIMD/MLFF training runs at various temperatures.

Is using a strictly fixed Gaussian smearing (ISMEAR = 0, SIGMA = 0.020 eV) a valid and physically acceptable compromise for training a metallic MLFF across different thermodynamic temperatures?

Any guidance on the standard best practices for this workflow would be greatly appreciated.

Thanks
Dominic

[zahedzx]

Hi Dominic,

Yes, your approach is reasonable. For MLFF training, consistency is usually more important than matching the electronic smearing to the ionic temperature. Using a fixed smearing scheme and a fixed SIGMA across all configurations ensures that the MLFF is trained on a single, consistent DFT energy surface.

For AIMD and MLFF dataset generation, ISMEAR = 0 with a fixed SIGMA is a common and sensible choice for metals, since Methfessel-Paxton is generally better suited for static calculations than MD.

The main point is to verify that your chosen SIGMA = 0.020 eV provides stable SCF convergence and smooth forces. If convergence becomes problematic, you may need to increase SIGMA slightly.

In summary, using ISMEAR = 0 and a fixed SIGMA = 0.020 eV for both the initial calculations and all subsequent AIMD/MLFF training runs is a valid and physically consistent strategy.

Best,
Zahed

[dominic_varghese]

Hi Zahed,

Thanks for the clarification.

I have a question regarding the workflow for training VASP MLFFs on metals. Because these calculations often hit the cluster walltime limit before completing the target number of MD steps, is it valid to restart the simulation using the final CONTCAR as the new POSCAR and loading the previously generated ML_ABN file? My goal is to ensure the model retains its previously learned configurations so the restarted run only triggers ab initio DFT steps for newly encountered, structurally distinct phases, thereby maximizing data generation efficiency.

Thanks
Dominic

[zahedzx]

Hi Dominic,

Yes, that is the recommended workflow. You can restart from the final CONTCAR (renamed to POSCAR) and reuse the previously generated ML_ABN by copying or renaming it to ML_AB. As described in the VASP Wiki, an ML_ABN file produced in ML_MODE = train or select can be reused as input for another ML_MODE = train or select calculation. This allows the model to retain the previously learned configurations and continue adding only new ones as needed.

Best,
Zahed