\(\renewcommand{\AA}{\text{Å}}\)

pair_style mliap command

Accelerator Variants: mliap/kk

Syntax

pair_style mliap ... keyword values ...
  • one or two keyword/value pairs must be appended

  • keyword = model or descriptor or unified

    model values = style filename
      style = linear or quadratic or nn or mliappy
      filename = name of file containing model definitions
    descriptor values = style filename
      style = sna or so3 or ace
      filename = name of file containing descriptor definitions
    unified values = filename ghostneigh_flag
      filename = name of file containing serialized unified Python object
      ghostneigh_flag = 0/1 to turn off/on inclusion of ghost neighbors in neighbors list

Examples

pair_style mliap model linear InP.mliap.model descriptor sna InP.mliap.descriptor
pair_style mliap model quadratic W.mliap.model descriptor sna W.mliap.descriptor
pair_style mliap model nn Si.nn.mliap.model descriptor so3 Si.nn.mliap.descriptor
pair_style mliap model mliappy ACE_NN_Pytorch.pt descriptor ace ccs_single_element.yace
pair_style mliap unified mliap_unified_lj_Ar.pkl 0
pair_coeff * * In P

Description

Pair style mliap provides a general interface to families of machine-learning interatomic potentials. It allows 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. The pair style currently supports sna, so3 and ace descriptor styles, but it is straightforward to add new descriptor styles. By using the unified keyword, it is possible to define a Python model that combines functionalities of both model and descriptor.

The SNAP descriptor style sna is the same as that used by pair_style snap, including the linear, quadratic, and chem variants. The available models are linear, quadratic, nn, and mliappy. The mliappy style can be used to couple python models, e.g. PyTorch neural network energy models, and requires building LAMMPS with the PYTHON package (see below). In order to train a model, it is useful to know the gradient or derivative of energy, force, and stress w.r.t. model parameters. This information can be accessed using the related compute mliap command.

Added in version 2Jun2022.

The descriptor style so3 is a descriptor that is derived from the the smooth SO(3) power spectrum with the explicit inclusion of a radial basis (Bartok) and (Zagaceta). The available models are linear and nn.

Added in version 17Apr2024.

The descriptor style ace is a class of highly general atomic descriptors, atomic cluster expansion descriptors (ACE) from (Drautz), that include a radial basis, an angular basis, and bases for other variables (such as chemical species) if relevant. In descriptor style ace, the ace descriptors may be defined up to an arbitrary body order. This descriptor style is the same as that used in pair_style pace and compute pace. The available models with ace in ML-IAP are linear and mliappy. The ace descriptors and models require building LAMMPS with the ML-PACE package (see below). The mliappy model style may be used with ace descriptors, but it requires that LAMMPS is also built with the PYTHON package. As with other model styles, the mliappy model style can be used to couple arbitrary python models that use the ace descriptors such as Pytorch NNs. Note that ALL mliap model styles with ace descriptors require that descriptors and hyperparameters are supplied in a .yace or .ace file, similar to compute pace.

The pair_style mliap command must be followed by two keywords model and descriptor in either order, or the one keyword unified. A single pair_coeff command is also required. The first 2 arguments must be * * so as to span all LAMMPS atom types. This is followed by a list of N arguments that specify the mapping of MLIAP element names to LAMMPS atom types, where N is the number of LAMMPS atom types.

The model keyword is followed by the model style. This is followed by a single argument specifying the model filename containing the parameters for a set of elements. The model filename usually ends in the .mliap.model extension. It may contain parameters for many elements. The only requirement is that it contain at least those element names appearing in the pair_coeff command.

The top of the model file can contain any number of blank and comment lines (start with #), but follows a strict format after that. The first non-blank non-comment line must contain two integers:

  • nelems = Number of elements

  • nparams = Number of parameters

When the model keyword is linear or quadratic, this is followed by one block for each of the nelem elements. Each block consists of nparams parameters, one per line. Note that this format is similar, but not identical to that used for the pair_style snap coefficient file. Specifically, the line containing the element weight and radius is omitted, since these are handled by the descriptor.

When the model keyword is nn (neural networks), the model file can contain blank and comment lines (start with #) anywhere. The second non-blank non-comment line must contain the string NET, followed by two integers:

  • ndescriptors = Number of descriptors

  • nlayers = Number of layers (including the hidden layers and the output layer)

and followed by a sequence of a string and an integer for each layer:

  • Activation function (linear, sigmoid, tanh or relu)

  • nnodes = Number of nodes

This is followed by one block for each of the nelem elements. Each block consists of scale0 minimum value, scale1 (maximum - minimum) value, in order to normalize the descriptors, followed by nparams parameters, including bias and weights of the model, starting with the first node of the first layer and so on, with a maximum of 30 values per line.

The detail of nn module implementation can be found at (Yanxon).

Notes on mliappy models

When the model keyword is mliappy, if the filename ends in ‘.pt’, or ‘.pth’, it will be loaded using pytorch; otherwise, it will be loaded as a pickle file. To load a model from memory (i.e. an existing python object), specify the filename as “LATER”, and then call lammps.mliap.load_model(model) from python before using the pair style. When using LAMMPS via the library mode, you will need to call lammps.mliappy.activate_mliappy(lmp) on the active LAMMPS object before the pair style is defined. This call locates and loads the mliap-specific python module that is built into LAMMPS.

The descriptor keyword is followed by a descriptor style, and additional arguments. Currently three descriptor styles are available: sna, so3, and ace.

  • sna indicates the bispectrum component descriptors used by the Spectral Neighbor Analysis Potential (SNAP) potentials of pair_style snap. 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.

  • so3 indicated the power spectrum component descriptors. A single additional argument specifies the descriptor filename containing the parameters and setting.

  • ace indicates the atomic cluster expansion (ACE) descriptors. A single additional argument specifies the filename containing parameters, settings, and definitions of the ace descriptors (through tabulated basis function indices and corresponding generalized Clebsch-Gordan coefficients) in the ctilde file format, e.g. in the potential file format with *.ace or *.yace extensions from pair_style pace. Note that unlike the potential file, the Clebsch-Gordan coefficients in the descriptor file supplied should NOT be multiplied by linear or square root embedding terms.

The SNAP descriptor file closely follows the format of the pair_style snap parameter file. The file can contain blank and comment lines (start with #) anywhere. Each non-blank non-comment line must contain one keyword/value pair. The required keywords are rcutfac and twojmax. There are many optional keywords that are described on the pair_style snap doc page. In addition, the SNAP descriptor file must contain the nelems, elems, radelems, and welems keywords. The nelems keyword specifies the number of elements provided in the other three keywords. The elems keyword is followed by a list of nelems element names that must include the element names appearing in the pair_coeff command, but can contain other names too. Similarly, the radelems and welems keywords are followed by lists of nelems numbers giving the element radius and element weight of each element. Obviously, the order in which the elements are listed must be consistent for all three keywords.

The SO3 descriptor file is similar to the SNAP descriptor except that it contains a few more arguments (e.g., nmax and alpha). The preparation of SO3 descriptor and model files can be done with the Pyxtal_FF package.

The ACE descriptor file differs from the SNAP and SO3 files. It more closely resembles the potential file format for linear or square-root embedding ACE potentials used in the pair_style pace. As noted above, the key difference is that the Clebsch-Gordan coefficients in the descriptor file with mliap descriptor ace are NOT multiplied by linear or square root embedding terms. In other words,the model is separated from the descriptor definitions and hyperparameters. In pair_style pace, they are combined. The ACE descriptor files required by mliap are generated automatically in FitSNAP during linear, pytorch, etc. ACE model fitting. Additional tools are provided there to prepare ace descriptor files and hyperparameters before model fitting. The ace descriptor files can also be extracted from ACE model fits in python-ace.. It is important to note that order of the types listed in pair_coeff must match the order of the elements/types listed in the ACE descriptor file for all mliap styles when using ace descriptors.

See the pair_coeff page for alternate ways to specify the path for these model and descriptor files.

Note

To significantly reduce SO3 descriptor/force calculation time, some properties are pre-computed and reused during the calculation. These can consume a significant amount of RAM for simulations of larger systems since their size depends on the total number of neighbors per MPI process.

Added in version 3Nov2022.

The unified keyword is followed by an argument specifying the filename containing the serialized unified Python object and the “ghostneigh” toggle (0/1) to disable/enable the construction of neighbors lists including neighbors of ghost atoms. If the filename ends in ‘.pt’, or ‘.pth’, it will be loaded using pytorch; otherwise, it will be loaded as a pickle file. If ghostneigh is enabled, it is recommended to set comm_modify cutoff manually, such as in the following example.

variable ninteractions equal 2
variable cutdist equal 7.5
variable skin equal 1.0
variable commcut equal (${ninteractions}*${cutdist})+${skin}
neighbor ${skin} bin
comm_modify cutoff ${commcut}

Note

To load a model from memory (i.e. an existing python object), call lammps.mliap.load_unified(unified) from python, and then specify the filename as “EXISTS”. When using LAMMPS via the library mode, you will need to call lammps.mliappy.activate_mliappy(lmp) on the active LAMMPS object before the pair style is defined. This call locates and loads the mliap-specific python module that is built into LAMMPS.


Styles with a gpu, intel, kk, omp, or opt suffix are functionally the same as the corresponding style without the suffix. They have been optimized to run faster, depending on your available hardware, as discussed on the Accelerator packages page. The accelerated styles take the same arguments and should produce the same results, except for round-off and precision issues.

These accelerated styles are part of the GPU, INTEL, KOKKOS, OPENMP, and OPT packages, respectively. They are only enabled if LAMMPS was built with those packages. See the Build package page for more info.

You can specify the accelerated styles explicitly in your input script by including their suffix, or you can use the -suffix command-line switch when you invoke LAMMPS, or you can use the suffix command in your input script.

See the Accelerator packages page for more instructions on how to use the accelerated styles effectively.


Mixing, shift, table, tail correction, restart, rRESPA info

For atom type pairs I,J and I != J, where types I and J correspond to two different element types, mixing is performed by LAMMPS with user-specifiable parameters as described above. You never need to specify a pair_coeff command with I != J arguments for this style.

This pair style does not support the pair_modify shift, table, and tail options.

This pair style does not write its information to binary restart files, since it is stored in potential files. Thus, you need to re-specify the pair_style and pair_coeff commands in an input script that reads a restart file.

This pair style can only be used via the pair keyword of the run_style respa command. It does not support the inner, middle, outer keywords.


Restrictions

This pair style 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 requires building LAMMPS with the PYTHON package. The ace descriptor requires building LAMMPS with the ML-PACE package. See the Build package page for more info. Note that pair_mliap/kk acceleration will not invoke the kk accelerated variants of SNAP or ACE descriptors.

Default

none


(Bartok2013) Bartok, Kondor, Csanyi, Phys Rev B, 87, 184115 (2013).

(Zagaceta2020) Zagaceta, Yanxon, Zhu, J Appl Phys, 128, 045113 (2020).

(Yanxon2020) Yanxon, Zagaceta, Tang, Matteson, Zhu, Mach. Learn.: Sci. Technol. 2, 027001 (2020).