"Conserved En." is increasing linearly with simulation time - Is this normal?

GROMACS version:2020.3
GROMACS modification: No

My aim is to compute the self-diffusivity of methanol in zeolite Y.
To achieve this, the first task is to equilibrate the system. To assess whether the system is equilibrated or not, I plot various quantities as a function of simulation time. These quantities are:
Potential energy, Kinetic energy, Total energy, Conserved energy, Temperature, and Pressure.
I get these quantities using the gmx energy command.

In spite of its name, the “conserved energy” is increasing linearly with simulation time (see attached plot). Is this a normal behavior?

I must admit I am not familiar with the quantity “conserved energy”. What is it describing?
Thanks a lot in advance for your help.


Dear Dr. Lemkul,

Thanks for your help.
Unfortunately, I am not able to understand the information in the link. Is there not any review article or text book that explains what this quantity is and/or its expected behavior?


Maybe I should add that I am doing an NVT simulation, and I am using the Nose-Hoover thermostat. I also attach the mdp file in case it contains relevant information.

Torstein Fjermestadminim.mdp (1.7 KB)

Pull out the old web search engine:

This answer on the mailing list from 2013 is incorrect.

The conserved energy quantity should be conserved analytically. However, discretization as well as approximations lead to small integration errors causing a linear drift in most cases. How large this drift can be be and what one can tolerate depends on the system and is often difficult to say.

I don’t know the size of your system, so I can’t say what reasonable for your case.

My system contains 4800 atoms.
The length of the simulation is 40 ns.
The total drift of the conserved energy is 1.11767e+06 kJ/mol

Normalizing, the drift of the conserved energy per ns and atoms becomes:
5.82 kJ mol-1 ns-1 atom-1

Does this look like a reasonable drift of the conserved energy?

A late answer: yes, that is a reasonable order of magnitude.

@hess , thanks a lot for your answer.
I now realize that it would also be nice to know how I can assess whether the drift is reasonable in a general case. You say above that it is often difficult to say, but is it not possible to make crude approximations based on knowledge available in the literature?

Are there any works I can refer to in order to justify that the drift is reasonable?
Thanks once again :)

Now your drift is nearly the same as the tolerance for the Verlet buffer for the pair list. You can’t expect it to be much smaller than that, unless you change that mdp option. But that mdp option was set to that value because constraints, in single precision, often give about an equal amount of drift. So for standard atomistic simulations this is reasonable.