[MMTK] Auto-discard notification

Sanzo Miyazawa miyazawa at smlab.sci.gunma-u.ac.jp
Thu Jun 22 11:51:07 CEST 2006

>From hinsen at cnrs-orleans.fr Thu Jun 22 16:04:45 2006
>The alphabetic order leads to a unique choice only when all three  
>atom types are different, which is not always the case. If two or  
>more atom types are identical, the choice is practically random, and  
>might well depend on the configuration of the computer. 
>Any program  
>implementing the Amber force field should have the same problem,  
>though those that (like Amber itself) define a system with a rigid  
>atom order have a better chance of getting the same result on all  
>computers - at the price of having the energy depend on an arbitrary  
>choice in the system definition (the atom indices).

>The only proper solution would be to make the energy dependent on  
>nothing but physical parameters. For improper dihedrals, the only way  
>I can see is to use the average of all three improper dihedral  
>choices. However, I hesitate to do this as it is in contradiction  
>with the Amber force field specification.
>Note that the three choices lead to very similar energies because the  
>three bond lengths are quite similar. The amibiguity in the  
>definition probably has no consequences for the interpretation of  
>simulation results.

I was very surprised to hear the fact above.

Although I cannot understand why the order of improper DH depends on a computer,
if so, it can be explained why different runs on different computers
and even on the same computer yield completely different MD trajaectories.

As you note, the difference is small for energy.
In the case of 1L2Y.pdb with Amber99ForceField, one result for energy was

{'electrostatic/pair sum': -2173.6696173251594, 'cosine dihedral angle': 750.789784524413
34, 'harmonic bond': 51.816364296890683, 'electrostatic': -2173.6696173251594, 'harmonic
bond angle': 163.46003875492255, 'Lennard-Jones': -238.79780891259904}

and the other result was

{'electrostatic/pair sum': -2173.6696173251594, 'cosine dihedral angle': 751.125460871967
33, 'harmonic bond': 51.816364296890683, 'electrostatic': -2173.6696173251594, 'harmonic
bond angle': 163.46003875492255, 'Lennard-Jones': -238.79780891259904}

Therefore in this case, the difference of the total energy is less than 0.03 %.
Also probably the difference on the energy gradient would be very small too.
However, evern though the difference is small, any difference on the
energy gradient can lead the system to the completely different and 
distant local minimum.

For me, reproducibility is more important than keeping a compatibility (?)
with the Amber's strange interpretation of the improper DHA.

Thank you very much for your help.

More information about the mmtk mailing list