I checked back to some of my previous simulations with 0.005\tau. In some simulations, the conservation of energy is really good for example single atoms EMD. Some have weak drift, i.e. binary mixing of two atom species. But for velocity-Verlet integrator, I believe 0.005\tau is commonly used time step value. Do you suggest if I use a smaller time step, it may be able to prevent the energy drift in RNEMD?