Broken bonds or deformation of molecules in long run using ML_FF after refit

[suojiang_zhang1]

Hi,
Following my previous post.
I used the refited ml_ff to run the MD simulation in a larger system and a longer time (200ps).
I think the trained FF is very accurate with a low ERR. But I found in this simulation process many molecules occur to the broken bond or deformation.
in this case, do I continue to train the FF? or is there another better approach to improve the case?

[ferenc_karsai]

Ok there is no recipe what to do here. It's usually a science question to obtain the correct fitting of a material and later obtain the right physics in a production run.

I can only give some general thoughts here:
The "ERR" in the ML_LOGFILE only provides errors on the training data. You can have a very good trainings error, but if your target structures in your production run are out of scope from your training data the force field will not be accurate or even completely fail.
This is most likely the case in your calculation.
We also saw in some calculations of liquids that they can be unstable if the on-the-fly training MD was not run long enough. There seemingly doesn't happen anything special if you run them much longer, but the force-fields get much more stable.
So you should try to run on-the-fly training much longer (for liquids it was several hundreds of picoseconds).

Of course another problem could be that your MD setup is not ok, but you haven't posted anything according to the forum guidelines.