The current reax/reaxc implementation uses CG minimization when doing charge equilibration (qeq).
I am just wondering if anyone is implementing an alternative
qeq method such as SQE, QTIPE or some other variant within LAMMPS reax/reaxc.
We are starting to create a more generalized charge equilibration
module, for use with other potentials. We might add an option
to use different kinds of minimizers with that, but please give
us some compelling arguments for preferring one over the other?
I believe the current reaxc implementation now uses a separate fix to perform qeq and is therefore decoupled from the covalent terms in the pair style.
Someone named “Pablo Valentini” ask a similar question back in July 24, 2009.
He wanted to know if some other qeq methods such as QTPIE (Sorry for the previous wrong abbreviation) have been implemented in LAMMPS for ReaxFF other than EEM/CG minimization.
My question is similar to Pablo’s. I am just wondering if anyone is currently working on such qeq for ReaxFF/reaxc at the moment.
I believe the qeq’s are specific to the potential. COMB for example uses extended Lagrangian and it’s uniquely embeded in the potential implementation. I am not sure if CG minimization would produce similar result when applied to systems using COMB.
Fix qeq/reax is indeed a separate fix in terms of its implementation.
Since the covalent terms of ReaxFF are charge independent, fix
qeq/reax does not need to take information from the pair_style, hence
by your definition "decoupled". However, this is the characteristics
of ReaxFF and has very little to do with the charge solving minimizer.
Any minimizer with pair_style reax/c will be "separate and decoupled"
from the pair_style.
QEq, QTPIE and fluc-q are variable charge models. ReaxFF and COMB
both use QEq (with some differences in details).
CG solver and extended Lagrangian are charge solving methods, or
minimizers; ReaxFF uses CG solver to minimize charge-dependent
energies E(q), COMB and fluc-q both use extended Lagrangian to
minimize charge forces (derivative of E(q) wrt charge). One can
indeed modify pair_style comb to use CG solver, but I am not sure it
will reproduce the exact same answers either.
QTPIE seems to add a distance-dependent function to penalize
long-range charge transfer, which is similar in some sense to the
taper function used in CG solver of fix qeq/reax. I agree with Steve
that a strong and compelling argument is needed for preferring QTPIE
over QEq. Please feel free to correct me if my understanding of QTPIE
Thanks for the insight. As you mentioned, QTPIE does penalize long range charge transfer. The fomulation also allows systems to have the correct charge behaviour when bond breaking occurs. However, the taper function in ReaxFF is meant to cut off Coulomb interactions so that no Ewald summation would be necessary.
ReaxFF uses CQ minimization to re-distribute charges every time step. Therefore charge transfer always occurs irrespective of the atomic or molecular position. I have done a test with ReaxxFF using FCC Cu metal block placed 100 Angstroms away from a CuO block in the same simulation box. After, running a simple NVE MD for long enough time (10 ns), the whole Cu metal block became charged, with some atoms even carrying negative charges. Both system eventually carry an effective charge, positive or negative, but the overall system is charge neutral.
Someone recently reported ( http://lammps.sandia.gov/threads/msg28994.html ) getting negative charges on Al atoms in Al2O3 system when using ReaxFF.
The current qeq using CG minimization as at now will always result in charge transfer even when two charge systems are at infinity. CG just bundle everyone together and perform energy minimization. If the mathematical solution require that a metal carries a negative charge to minimize energy, CG method does not give a damn as long as the energy is minimized and all charges sums up to zero.
Methods such as SQE, QTPIE and other variants make sure that these physical phenomenon are not violated. There is a new AIREBO potential for C-H-O that implements a modified form of SQE with better results and avoid this unphysical behavior.
So the summary is that, as it stand right now, the qeq in reaxx does not suit my application, that is why I believe alternative methods are in order. We are devising our own solution, but I just wanted to know if someone is doing similar work in LAMMPS for ReaxFF at the moment.
This is very good analysis and it does provide a sound argument.
Perhaps when we are in the process of creating a more generalized
variable charge equilibration method we will consider including this
as an optional choice. Many thanks.