fix qeq/reaxff command

Accelerator Variants: qeq/reaxff/kk, qeq/reaxff/omp

Syntax

fix ID group-ID qeq/reaxff Nevery cutlo cuthi tolerance params args

  • ID, group-ID are documented in fix command

  • qeq/reaxff = style name of this fix command

  • Nevery = perform QEq every this many steps

  • cutlo,cuthi = lo and hi cutoff for Taper radius

  • tolerance = precision to which charges will be equilibrated

  • params = reaxff or a filename

  • one or more keywords or keyword/value pairs may be appended

    keyword = dual or maxiter or nowarn or matfree or xlmd or xlcg or xldamp or xlkappa or xlguard
  dual = process S and T matrix in parallel (only for qeq/reaxff/omp)
  maxiter N = limit the number of iterations to N
  nowarn = do not print a warning message if the maximum number of iterations was reached
  matfree = use a matrix-free approach for applying the H matrix (only for qeq/reaxff/kk)
  xlmd value = yes or no = whether or not to use the extended-Lagrangian propagation of the charges
  xlcg value = number of conjugate gradient iterations per step (requires xlmd yes)
  xldamp value = order of the dissipation term for the auxiliary variables, 0 or 3 to 9 (requires xlmd yes)
  xlkappa value = coupling constant \(\kappa = \omega^2 \delta t^2\) of the undamped propagation (requires xlmd yes)
  xlguard value = relative residual of the truncated solves above which the extended-Lagrangian propagation is restarted, <= 0 disables (requires xlmd yes)

Examples

fix 1 all qeq/reaxff 1 0.0 10.0 1.0e-6 reaxff
fix 1 all qeq/reaxff 1 0.0 10.0 1.0e-6 param.qeq maxiter 500
fix 1 all qeq/reaxff 1 0.0 10.0 1.0e-6 reaxff xlmd yes

Description

Perform the charge equilibration (QEq) method as described in (Rappe and Goddard) and formulated in (Nakano). It is typically used in conjunction with the ReaxFF force field model as implemented in the pair_style reaxff command, but it can be used with any potential in LAMMPS, so long as it defines and uses charges on each atom. The fix qeq/comb command should be used to perform charge equilibration with the COMB potential. For more technical details about the charge equilibration performed by fix qeq/reaxff, see the (Aktulga) paper.

The QEq method minimizes the electrostatic energy of the system by adjusting the partial charge on individual atoms based on interactions with their neighbors. It requires some parameters for each atom type. If the params setting above is the word “reaxff”, then these are extracted from the pair_style reaxff command and the ReaxFF force field file it reads in. If a file name is specified for params, then the parameters are taken from the specified file and the file must contain one line for each atom type. The latter form must be used when performing QEq with a non-ReaxFF potential. Each line should be formatted as follows:

itype chi eta gamma

where itype is the atom type from 1 to Ntypes, chi denotes the electronegativity in eV, eta denotes the self-Coulomb potential in eV, and gamma denotes the valence orbital exponent. Note that these 3 quantities are also in the ReaxFF potential file, except that eta is defined here as twice the eta value in the ReaxFF file. Note that unlike the rest of LAMMPS, the units of this fix are hard-coded to be A, eV, and electronic charge.

The optional dual keyword allows to perform the optimization of the S and T matrices in parallel. This is only supported for the qeq/reaxff/omp style. Otherwise they are processed separately. The qeq/reaxff/kk style always solves the S and T matrices in parallel.

The optional maxiter keyword allows changing the max number of iterations in the linear solver. The default value is 200.

The optional nowarn keyword silences the warning message printed when the maximum number of iterations was reached. This can be useful for comparing serial and parallel results where having the same fixed number of QEq iterations is desired, which can be achieved by using a very small tolerance and setting maxiter to the desired number of iterations.

Added in version TBD.

The optional xlmd keyword with the value yes enables the extended-Lagrangian propagation of the charges, which is explained in more detail for the fix qeq/shielded/xlmd command. Auxiliary per-atom variables are propagated time-reversibly alongside the atoms and provide the initial guess for the iterative solver, which then applies only a small, fixed number of conjugate gradient iterations (keyword xlcg, default 2) instead of iterating to tolerance. This makes the charge equilibration part of a time step about 1.5 to 3 times faster (see the estimates below), while the energy conservation usually remains close to that of fully converged solves. The xldamp and xlkappa keywords set the dissipation and the coupling of the auxiliary variables in the same way as for fix qeq/shielded/xlmd (default: dissipation of order 5). The xlguard keyword sets the threshold of a safeguard against occasional instabilities of the propagation, which is also explained there: when the relative residual of a truncated solve exceeds this value (default 1.0e-3), the solve is completed to tolerance, a warning is printed (unless the nowarn keyword is used), and the propagation is restarted from the converged charges. A value <= 0 disables the safeguard. At the beginning of every run, after the number of atoms changed, and during energy minimization, the charges are equilibrated to tolerance until a valid history of solutions from consecutive steps is available. The extended-Lagrangian propagation requires Nevery = 1. It is available for all variants of this fix and for the fix qtpie/reaxff and fix qeq/rel/reaxff commands, but not for fix acks2/reaxff. It is recommended to verify the energy conservation of a simulation against a run with fully converged charges.

The optional matfree keyword replaces the sequence of explicitly constructing the H matrix, then (repeatedly) applying it with a matrix-free approach where the H matrix is effectively regenerated each time it is applied. This trades performance for reduced memory requirements because it avoids the overheads of storing the matrix. This is only supported for the qeq/reaxff/kk style, with both full and half qeq neighbor lists supported.

Added in version TBD.

The following estimates for using the xlmd and matfree keywords are from the PETN crystal benchmark in the bench/POTENTIALS folder for a system of about 1M atoms. With xlmd yes, only 3 conjugate gradient iterations per time step are used instead of an average of about 12. This reduced the time per step by about 15% on 8 CPU cores, by about 20% on one GPU, and by 35% on four GPUs. The savings are larger on multiple GPUs, since every iteration requires communication between them. With the matfree keyword, the GPU memory needed was about a third smaller, so that up to 75% more atoms fit on one GPU, but the simulation took about 50% longer with GPUs and more than twice as much time on the CPU. Combined with xlmd yes, which applies the matrix only a few times per step, the simulation was still up to 35% faster on GPUs and about as fast on the CPU. The actual savings depend on the individual system, the tolerance, the hardware, and the number of MPI processes or GPUs.

Changed in version TBD.

If the fix group does not contain all atoms, only the charges of the atoms in the fix group are equilibrated. The charges of all other atoms are not changed, but they still contribute to the electrostatic potential at the atoms in the fix group, i.e. they act as fixed external charges. All accelerated variants of this fix follow the same convention.

Note

In order to solve the self-consistent equations for electronegativity equalization, LAMMPS imposes the additional constraint that all the charges in the fix group must add up to zero. The initial charge assignments should also satisfy this constraint. LAMMPS will print a warning if that is not the case.

Restart, fix_modify, output, run start/stop, minimize info

No information about this fix is written to binary restart files. This fix computes a global scalar (the number of iterations) for access by various output commands. No parameter of this fix can be used with the start/stop keywords of the run command.

This fix is invoked during energy minimization.

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.

Restrictions

This fix is part of the REAXFF package. It is only enabled if LAMMPS was built with that package. See the Build package page for more info.

This fix does not correctly handle interactions involving multiple periodic images of the same atom. Hence, it should not be used for periodic cell dimensions smaller than the non-bonded cutoff radius, which is typically \(10~\AA\) for ReaxFF simulations.

This fix may be used in combination with fix efield and will apply the external electric field during charge equilibration, but there may be only one fix efield instance used and the electric field vector may only have components in non-periodic directions. Equal-style variables can be used for electric field vector components without any further settings. Atom-style variables can be used for spatially-varying electric field vector components, but the resulting electric potential must be specified as an atom-style variable using the potential keyword for fix efield.

Default

maxiter 200, xlmd no, xlcg 2, xldamp 5, xlguard 1.0e-3. The last three settings only apply for xlmd yes.

(Rappe) Rappe and Goddard III, Journal of Physical Chemistry, 95, 3358-3363 (1991).

(Nakano) Nakano, Computer Physics Communications, 104, 59-69 (1997).

(Aktulga) Aktulga, Fogarty, Pandit, Grama, Parallel Computing, 38, 245-259 (2012).