pair_style hdnnp command¶
pair_style hdnnp cutoff keyword value ...
cutoff = short-range cutoff of HDNNP (maximum symmetry function cutoff radius)
zero or more keyword/value pairs may be appended
keyword = dir or showew or showewsum or maxew or resetew or cflength or cfenergy
value depends on the preceding keyword:
dir value = directory directory = Path to HDNNP configuration files showew value = yes or no showewsum value = summary summary = Write EW summary every this many timesteps (0 turns summary off) maxew value = threshold threshold = Maximum number of EWs allowed resetew value = yes or no cflength value = length length = Length unit conversion factor cfenergy value = energy energy = Energy unit conversion factor
pair_style hdnnp 6.35 showew yes showewsum 100 maxew 1000 resetew yes cflength 1.8897261328 cfenergy 0.0367493254 pair_coeff * * H O pair_style hdnnp 6.01 dir "./" showewsum 10000 pair_coeff * * S Cu NULL Cu
This pair style adds an interaction based on the high-dimensional neural network potential (HDNNP) method as presented in (Behler and Parrinello 2007). HDNNPs are machine learning potentials which require careful training of neural networks prior to application in MD simulations. The pair style uses an interface to the n2p2 library (Singraber, Behler and Dellago 2019) which is available on Github here. Please see the n2p2 documentation for further details. n2p2 (and hence this pair style) is compatible with neural network potentials trained with its own tools (Singraber et al 2019) and with RuNNer.
Only a single pair_coeff command with two asterisk wild-cards is used with this pair style. Its additional arguments define the mapping of LAMMPS atom types to n2p2 elements.
pair_coeff * * H O
In the above example LAMMPS types 1 and 2 are mapped to the elements “H” and “O” in n2p2, respectively. Multiple types may map to the same element, or some types may not be mapped at all. For example, if the LAMMPS simulation has four atom types, the command
pair_coeff * * H H O NULL
maps atom types 1 and 2 to the element “H”, type 3 to “O” and type 4 is not mapped (indicated by NULL). Atoms mapped to NULL are ignored by the HDNNP calculation, i.e. they do not contribute in any way to the evaluation of HDNNP energies and forces. This may be useful in a setup with hybrid pair styles.
The mandatory pair style argument cutoff must match the short-range cutoff radius of the HDNNP. This corresponds to the maximum cutoff radius of all symmetry functions (the atomic environment descriptors of HDNNPs) used.
The cutoff must be given in LAMMPS length units, even if the neural network potential has been trained using a different unit system (see remarks about the cflength and cfenergy keywords below for details).
The numeric value may be slightly larger than the actual maximum symmetry function cutoff radius (to account for rounding errors when converting units), but must not be smaller.
Use the dir keyword to specify the directory containing the HDNNP configuration
files. The directory must contain
input.nn with neural network and symmetry
scaling.data with symmetry function scaling data and
weights.???.data with weight parameters for each element.
The keyword showew can be used to turn on/off the display of extrapolation
warnings (EWs) which are issued whenever a symmetry function value is out of
bounds defined by minimum/maximum values in
scaling.data. An extrapolation
warning may look like this:
### NNP EXTRAPOLATION WARNING ### STRUCTURE: 0 ATOM: 119 ELEMENT: Cu SYMFUNC: 32 TYPE: 3 VALUE: 2.166E-02 MIN: 2.003E-05 MAX: 1.756E-02
stating that the value 2.166E-02 of symmetry function 32 of type 3 (Narrow Angular symmetry function), element Cu (see the log file for a symmetry function listing) was out
of bounds (maximum in
scaling.data is 1.756E-02) for atom 119. Here, the
atom index refers to the LAMMPS tag (global index) and the structure index is
used to print out the MPI rank the atom belongs to.
The showew keyword should only be set to yes for debugging purposes. Extrapolation warnings may add lots of overhead as they are communicated each timestep. Also, if the simulation is run in a region where the HDNNP was not correctly trained, lots of extrapolation warnings may clog log files and the console. In a production run use showewsum instead.
The keyword showewsum can be used to get an overview of extrapolation warnings occurring during an MD simulation. The argument specifies the interval at which extrapolation warning summaries are displayed and logged. An EW summary may look like this:
### NNP EW SUMMARY ### TS: 100 EW 11 EWPERSTEP 1.100E-01
Here, at timestep 100 the occurrence of 11 extrapolation warnings since the last summary is reported, which corresponds to an EW rate of 0.11 per timestep. Setting showewsum to 0 deactivates the EW summaries.
A maximum number of allowed extrapolation warnings can be specified with the maxew keyword. If the number of EWs exceeds the maxew argument the simulation is stopped. Note however that this is merely an approximate threshold since the check is only performed at the end of each timestep and each MPI process counts individually to minimize communication overhead.
The keyword resetew alters the behavior of the above mentioned maxew threshold. If resetew is set to yes the threshold is applied on a per-timestep basis and the internal EW counters are reset at the beginning of each timestep. With resetew set to no the counters accumulate EWs along the whole trajectory.
If the training of a neural network potential has been performed with different physical units for length and energy than those set in LAMMPS, it is still possible to use the potential when the unit conversion factors are provided via the cflength and cfenergy keywords. If for example, the HDNNP was parameterized with Bohr and Hartree training data and symmetry function parameters (i.e. distances and energies in “input.nn” are given in Bohr and Hartree) but LAMMPS is set to use metal units (Angstrom and eV) the correct conversion factors are:
cflength 1.8897261328 cfenergy 0.0367493254
Thus, arguments of cflength and cfenergy are the multiplicative factors required to convert lengths and energies given in LAMMPS units to respective quantities in native HDNNP units (1 Angstrom = 1.8897261328 Bohr, 1 eV = 0.0367493254 Hartree).
Mixing, shift, table, tail correction, restart, rRESPA info¶
This style does not support mixing. The pair_coeff command should only be invoked with asterisk wild cards (see above).
This style does not support the pair_modify shift, table, and tail options.
This style does not write information to binary restart files. Thus, you need to re-specify the pair_style and pair_coeff commands in an input script that reads a restart file.
This style can only be used via the pair keyword of the run_style respa command. It does not support the inner, middle, outer keywords.
This pair style is part of the ML-HDNNP package. It is only enabled if LAMMPS was built with that package. See the Build package doc page for more info.
Please report bugs and feature requests to the n2p2 GitHub issue page.
The default options are dir = “hdnnp/”, showew = yes, showewsum = 0, maxew = 0, resetew = no, cflength = 1.0, cfenergy = 1.0.
(Behler and Parrinello 2007) Behler, J.; Parrinello, M. Generalized Neural-Network Representation of High-Dimensional Potential-Energy Surfaces. Phys. Rev. Lett. 2007, 98 (14), 146401.
(Singraber, Behler and Dellago 2019) Singraber, A.; Behler, J.; Dellago, C. Library-Based LAMMPS Implementation of High-Dimensional Neural Network Potentials. J. Chem. Theory Comput. 2019, 15 (3), 1827-1840
(Singraber et al 2019) Singraber, A.; Morawietz, T.; Behler, J.; Dellago, C. Parallel Multistream Training of High-Dimensional Neural Network Potentials. J. Chem. Theory Comput. 2019, 15 (5), 3075-3092.