Bonded atoms far aprt in Coarse-grained simulation

Dear lammps user,

I am doing a coarse-grained simulation about a single polymer chain using LAMMPS (1 Mar 2016-ICMS). Firstly, I ran a NVT simulation with LJ potential as pair potential, Morse potential as bond potential and cosine/squared potential as angle potential and then it is all right. However, when I added a dihedral potential to the coarse-grained model, the bonded atoms of the polymer chain are far apart from each other and blowing away from the box. I also try the latest version of LAMMPS to run the simulation but the blowing still occurs.

The input script, data file and log file all are attached. The log file only includes the 2000 timesteps in which some atoms were blow away from the box.

It is werid that the bonded atoms are far apart from each other because the bonded interaction is strong enough. I also plot the curve of Kinetic energy and potential energy of the system when the boned atoms are far apart and they are showed below.
In the above two picture, it is strange that the kinetic energy increases suddenly prior to the potential energy. It is usual that the large force from the steep potential gradient results in the incresing of velocity and then kinetic energy. Is it possible that the calculation of velocity of a certain atom when simulating is incorrect? Or can anyone explains this sitiuation?

Any advice is welcome. Thanks for your attention and efforts in advance. (2.02 KB) (1.37 KB)

sd2.log (225 KB)