model
This keyword selects the model used for training or prediction.
Syntax
model <name>
Parameters
The names match the model field in nep.txt. If no model-selection
keyword is given, the default is nep.
Name |
Model |
|---|---|
|
Standard energy NEP |
|
TNEP dipole |
|
TNEP polarizability |
|
Temperature-dependent NEP |
|
Spin-dependent energy NEP |
|
Charge NEP with real- and reciprocal-space electrostatics |
|
Charge NEP with reciprocal-space electrostatics |
|
|
|
NEP with learned long-range vdW |
|
|
Specify model at most once.
For model nep_spin, spin_version defaults to 1.
It is an error to combine spin_version
with any other explicit model.
See flip_charge for the restart option.
Output
None.
Examples
model nep
Notes
Long-range models
The models qnep1, qnep2, and qnep3 [Fan2026] predict environment-dependent atomic charges in addition to the local NEP energy.
The qnep1 and qnep2 models follow the approaches of Song et al. [Song2024b] and Cheng [Cheng2025], respectively.
For periodic systems, their electrostatic contributions are
Here, \(E_{\mathrm{real}}\) and \(E_{\mathrm{k}}\) are the real- and reciprocal-space Ewald terms.
The correction \(E_{\mathrm{self}}=-k_{\mathrm e}\alpha\sum_i q_i^2/\sqrt{\pi}\) removes the charge self interaction, where \(k_{\mathrm e}\) is the Coulomb prefactor and \(\alpha\) is the Ewald splitting parameter.
In qnep2 and qnep3, the local NEP energy represents the remaining short-range interactions.
The nep_vdw model predicts environment-dependent dispersion coefficients and adds the long-range part of the attractive \(r^{-6}\) interaction.
For periodic systems, this part is evaluated in reciprocal space with its self interaction removed; the local NEP energy describes the short-range interactions.
The qnep3_vdw model combines this dispersion contribution with the charge contribution used by qnep3.
Training data
Use the energy, force, and optional virial or stress labels described in train.xyz and test.xyz.
Atomic charge and dispersion-coefficient labels are not required.
For the charge models, charge= specifies the total charge of each structure in units of e and defaults to zero.
Optional bec:R:9 labels provide target Born effective charges.
Fitting Born effective charges is usually appropriate for a single material in a single phase, since a constant high-frequency dielectric constant is assumed for the whole training dataset.
The weights of the total-charge and Born-effective-charge loss terms are set by lambda_q and lambda_z, respectively.
Boundary conditions and MD use
For training and prediction with nep, all five long-range models support fully periodic (pbc="T T T") and fully nonperiodic (pbc="F F F") structures.
The default is fully periodic; mixed boundary conditions are not supported.
For nonperiodic structures, the long-range contribution is evaluated by direct summation without periodic images.
For MD with gpumd, these models require periodic boundaries in all three directions and use a single GPU.
The charge models enforce zero total charge during MD.
Load the trained model in run.in using the potential keyword:
potential nep.txt
The model is selected from the contents of nep.txt.
The kspace keyword selects PPPM (the default) or Ewald summation for the reciprocal-space charge and dispersion contributions.