compute mliap command


compute ID group-ID mliap ... keyword values ...
  • ID, group-ID are documented in compute command

  • mliap = style name of this compute command

  • two or more keyword/value pairs must be appended

  • keyword = model or descriptor or gradgradflag

    model values = style
      style = linear or quadratic or mliappy
    descriptor values = style filename
      style = sna
      filename = name of file containing descriptor definitions
    gradgradflag value = 0/1
      toggle gradgrad method for force gradient


compute mliap model linear descriptor sna Ta06A.mliap.descriptor


Compute style mliap provides a general interface to the gradient of machine-learning interatomic potentials with respect to model parameters. It is used primarily for calculating the gradient of energy, force, and stress components with respect to model parameters, which is useful when training mliap pair_style models to match target data. It provides separate definitions of the interatomic potential functional form (model) and the geometric quantities that characterize the atomic positions (descriptor). By defining model and descriptor separately, it is possible to use many different models with a given descriptor, or many different descriptors with a given model. Currently, the compute supports just two models, linear and quadratic, and one descriptor, sna, the SNAP descriptor used by pair_style snap, including the linear, quadratic, and chem variants. Work is currently underway to extend the interface to handle neural network energy models, and it is also straightforward to add new descriptor styles.

The compute mliap command must be followed by two keywords model and descriptor in either order.

The model keyword is followed by the model style (linear, quadratic or mliappy). The mliappy model is only available if LAMMPS is built with the mliappy Python module. There are specific installation instructions for that module.

The descriptor keyword is followed by a descriptor style, and additional arguments. The compute currently supports two descriptor styles sna and so3, but it is is straightforward to add additional descriptor styles. The SNAP descriptor style sna is the same as that used by pair_style snap, including the linear, quadratic, and chem variants. A single additional argument specifies the descriptor filename containing the parameters and setting used by the SNAP descriptor. The descriptor filename usually ends in the .mliap.descriptor extension. The format of this file is identical to the descriptor file in the pair_style mliap, and is described in detail there.


The number of LAMMPS atom types (and the value of nelems in the model) must match the value of nelems in the descriptor file.

Compute mliap calculates a global array containing gradient information. The number of columns in the array is nelems \(\times\) nparams + 1. The first row of the array contain the derivative of potential energy with respect to. to each parameter and each element. The last six rows of the array contain the corresponding derivatives of the virial stress tensor, listed in Voigt notation: pxx, pyy, pzz, pyz, pxz, and pxy. In between the energy and stress rows are the \(3N\) rows containing the derivatives of the force components. See section below on output for a detailed description of how rows and columns are ordered.

The element in the last column of each row contains the potential energy, force, or stress, according to the row. These quantities correspond to the user-specified reference potential that must be subtracted from the target data when training a model. The potential energy calculation uses the built in compute thermo_pe. The stress calculation uses a compute called mliap_press that is automatically created behind the scenes, according to the following command:

compute mliap_press all pressure NULL virial

See section below on output for a detailed explanation of the data layout in the global array.

The optional keyword gradgradflag controls how the force gradient is calculated. A value of 1 requires that the model provide the matrix of double gradients of energy with respect to both parameters and descriptors. For the linear and quadratic models this matrix is sparse and so is easily calculated and stored. For other models, this matrix may be prohibitively expensive to calculate and store. A value of 0 requires that the descriptor provide the derivative of the descriptors with respect to the position of every neighbor atom. This is not optimal for linear and quadratic models, but may be a better choice for more complex models.

Atoms not in the group do not contribute to this compute. Neighbor atoms not in the group do not contribute to this compute. The neighbor list needed to compute this quantity is constructed each time the calculation is performed (i.e., each time a snapshot of atoms is dumped). Thus it can be inefficient to compute/dump this quantity too frequently.


If the user-specified reference potentials includes bonded and non-bonded pairwise interactions, then the settings of special_bonds command can remove pairwise interactions between atoms in the same bond, angle, or dihedral. This is the default setting for the special_bonds command, and means those pairwise interactions do not appear in the neighbor list. Because this fix uses the neighbor list, it also means those pairs will not be included in the calculation. The rerun command is not an option here, since the reference potential is required for the last column of the global array. A work-around is to prevent pairwise interactions from being removed by explicitly adding a tiny positive value for every pairwise interaction that would otherwise be set to zero in the special_bonds command.

Output info

Compute mliap evaluates a global array. The columns are arranged into nelems blocks, listed in order of element I. Each block contains one column for each of the nparams model parameters. A final column contains the corresponding energy, force component on an atom, or virial stress component. The rows of the array appear in the following order:

  • 1 row: Derivatives of potential energy with respect to each parameter of each element.

  • \(3N\) rows: Derivatives of force components; the x, y, and z components of the force on atom i appear in consecutive rows. The atoms are sorted based on atom ID.

  • 6 rows: Derivatives of the virial stress tensor with respect to each parameter of each element. The ordering of the rows follows Voigt notation: pxx, pyy, pzz, pyz, pxz, pxy.

These values can be accessed by any command that uses a global array from a compute as input. See the Howto output doc page for an overview of LAMMPS output options. To see how this command can be used within a Python workflow to train machine-learning interatomic potentials, see the examples in FitSNAP.


This compute is part of the ML-IAP package. It is only enabled if LAMMPS was built with that package. In addition, building LAMMPS with the ML-IAP package requires building LAMMPS with the ML-SNAP package. The mliappy model also requires building LAMMPS with the PYTHON package. See the Build package page for more info.


The keyword defaults are gradgradflag = 1