[lammps-users] about the accuracy and efficiency of REAXFF in LAMMPS

Hi Aidan:

Thanks for your help. Now I can run LAMMPS with REAX force field using C-H-O parameters. But I found some problems after I performed several MD simulations. The first thing is the low efficiency of REAX force field. It took more than 40 hours to carry out 2,000,000 MD run for the system of 80 carbon atoms at timestep of 0.02fs on a single processor. It’s really a long time for so small simulation system. I’m wondering if it is necessary to perform the MD simulation with REAXFF on multi processors. The second thing is the accuracy of REAXFF. I performed the NVT-MD simulation for the stability of an armchair CNT including 80 atoms at 1000K . I found the CNT started to be dissociated at 2,000,000 steps with the timestep of 0.02fs. I used other potentials such as AIREBO potential to do it, too, which showed the CNT is stable at 1000K. So, I’m wondering if REAXFF is not accurate for CNT systems or C-H-O REAX force field parameters are not fit for the CNT system. Aidan, do you have any suggestions? Have you ever tested the accuracy of REAXFF? For your reference, I attached the input file and the configuration of the CNT at 40ps. Your any suggestions would be highly appreciated.

Have a nice day,


in.CNT80-1000-40ps (624 Bytes)

ACNT80-1000-40ps v3.bmp (1.18 MB)

Your timing is pretty typical. To a surprising degree, the throughput on ReaxFF always lies close to 1 CPU-ms/atom-timestep. This is true across many platforms and applications. Normally I would suggest that you increase the processor count until you have less than 1000 atoms per processor, but you only have 80 atoms, so that won’t work. You will just have to be patient. In a few months, we may have a faster implementation of ReaxFF.

In order to examine the accuracy of the ReaxFF that you are using, you should really compare with a more accurate theory, such as DFT. You should also compare with relevant ReaxFF papers.

hi aidan and steve,

I have also encountered the same problem, i.e., ReaxFF potential is not stable. I have used a CNT equilibrated at 300K and run a NVE without controlling the temp. However, the temp fluctuate from 100 K to 600 K. Accordingly the pair potential changes a lot. I noticed that ReaxFF using real units and I carefully checked the consistency of using the unit and not found anything wrong. And this model works well for Airebo and Brenner potential. So I’m guessing there is sth wrong with the implement of Reax in lammps. It is still early to use it to do research… Hope the developer can solve this problem soon. We’re looking forward to using it in the near future. Thanks for your efforts.



This may be an issue with the particular version of ReaxFF in LAMMPS, which was not used for CNT.


My previous message could be misleading. Actually, I tested the simulation again. Now I find the CNT system is stable at 1000K. using free or fixed boundary conditions can result in the instability. But the correct periodic boundary condition is fine. I also test the NVE simulation of a CNT using periodic boundary condition. It is also stable. So, I think the boundary condition setting is very important for the REAX force field.


Hi, Lan,

What do you mean “stable”? not break up or crash the simulation? If you wish, I could send you my input script and test if the temperature fluctuation is very huge even at room temperature under periodic boundary condition.