lmp simulation of a disordered system

[J-You]

Hi all,

I trained a DP potential using the AIMD trajectories at liquid state. The train result seems good.

Then, I run lammps by this DP potential to simulate liquid structure. The initial configuration data is disordered, but the output contains large atomic clustering.

The version of deepmd is 1.3.3 gpu. AIMD box contains 200 atoms. Training data is sufficient.
I have simulated other system successfully. I think it is not caused by training parameter.

Have anyone ever been in a similar situation?

Thanks for your time and attention.

[njzjz]

A possible reason is your sel parameter is not enough. You can check it.
You can also print out the model deviation.

[343333333]

i used to train 64H2O model in dpmd , i meet the same problem like you . after i change my sel and it seems work well. A smaller sampling interval may also be effective

[J-You]

Thanks for your response. Do you mean that the `sel' is too small in my situation?
Could you please tell me more about how to change sampling interval, does it refers to batch size?

[343333333]

Thanks for your response. Do you mean that the `sel' is too small in my situation?
Could you please tell me more about how to change sampling interval, does it refers to batch size?

-its a odd thing to choose a good sel , i used sel of 96 which have bad performance on lammps , but 98 has a good performance . So , you should test a lot
-I mean the sampling interval t in lammps , water system performance better(less chance to bankrupt ) with the interval of 0.5fs than 1fs, because of the hydrogen.

[J-You]

Really appreciate for your suggestion. It does need a lot of test in my own training model.

[zhuhaotian519]

Hi, I had a similar problem with the DP model causing the MD trajectory to be messed up, how did you fix it later? Thank you!