[gmx-users] Conserved energy ("Conserved En.") in NVT simulation (Wade)

[Wade]

Dear Mark and Lu,
  Thank you very much for your responses. 
Based on your suggestions, I performed some tests:
1) I considered the influence of the time-step (with LINCS constraint), thermostat, and simulation precision in NVT simulation 
(tip4p water box, 300K, after 2ns NVE simulation). 
Below are results:
-----------
Time-step (fs)		thermostat (V or N)	 precision (S or D)	drift/ns = delt E/av E*100%
0.1		V		D		0.3
1		V		D		2
2		V		D		5.8


0.1		V		S		80
1		V		S		1.5
2		V		S		6.4


0.1		N		D		0.3
1		N		D		2.5
2		N		D		20


0.1		N		S		76
1		N		S		1.7
2		N		S		11.5
------------------
Above results shown that no matter the v-rescale or nose-hoover thermostat, the conserved Energy has nearly no drift (~0.3%/ns)
 when an very small time-step (0.1 fs) has been employed. But, even with 0.1fs step length, the integration can only be conserved in 
the double precision simulations.
For a commonly used time-step, such as 2 fs (with LINCS constraint), the drifts of the conserved energy are about 6-7% per ns. 
Such a drift might be not very significant in a short simulation. But, its accumulation in a long-time (e.g. several ns) simulation should be huge.


2) I checked the time-dependency of auxiliary dynamics purely, and found that it was the origination of the constant-rate drift of conserved energy. 
I also noticed that the exact solution of Eq.7 in bussi’s paper (Eq. A7) has been realized in gromacs to get the rescaling factor. 
But, I don’t know what the origination of the integration error is, and How can we deal with it.
As Lu said, does it comes from the summation of dK? If yes, how can we deal with the drift when we use an affordable time-step in a long-time simulation?


With best wishes,


Wade
‍