Our MD experts have looked at the details of this case. All cases where we observed that energy was not conserved in NVE, it was caused by a too large time step for the chosen potential, when it e.g. uses a spring that is too stiff, or it can also happen with ReaxFF. In the case here, we didn’t see any obvious reason, but also noticed that it is a slow drift which happens over 100s of ps. On this timescale it is not necessarily something unusual given that the numerical integration with a finite time step is an approximation and does not guarantee to conserve the energy 100% exactly. And your potential uses harmonic springs, which have fast vibrations, which would normally indicate a relatively fast time step to accurately conserve the energy.

So, the main suggestion is to lower the time step.