Hi,
thank you for your suggestion,
I can find the numbers from my ML_AB, and the numbers for H B C N F elements are 5000 1379 5000 3317 5000, respectively.
The temperatures were 200-600K in my train, and is 400K in my production run.
The pressure were 1 atm in my train by PSTRESS=0.001, and it was not changed in my production run.
the molecular bond broken and deformation after replicating the cell via ML_FF run
Moderators: Global Moderator, Moderator
-
suojiang_zhang1
- Newbie
- Posts: 44
- Joined: Tue Nov 19, 2019 4:15 am
-
jonathan_lahnsteiner2
- Global Moderator
- Posts: 215
- Joined: Fri Jul 01, 2022 2:17 pm
Re: the molecular bond broken and deformation after replicating the cell via ML_FF run
Dear suojiang_zhang1,
The values you are reporting sound reasonable. Did you also do a test set error
estimation at the simulation temperature that you are running at.
You would have to collect let's say 500 structures at 400K. You can use
structures from the MD runs you already have. But don't take structures which are
in the XDATCAR after each other. There should be some space between the extracted structures.
For the extracted structures you would have to make DFT calculations with the same settings
as when you trained the force field. Afterwards you would have to compute the root mean square
error between the DFT calculations and the MLFF calculations. You should compute the root mean square for the energy, the forces and the stress tensor.
You could also try to do a DFT MLFF comparison for the structures directly before your bond breaking occurs. In principle, if the error goes up the closer you get to the structure with broken bond the more certain it is that you have to explore more phase space during training.
Please let me know about your findings to discuss further.
All the best Jonathan
The values you are reporting sound reasonable. Did you also do a test set error
estimation at the simulation temperature that you are running at.
You would have to collect let's say 500 structures at 400K. You can use
structures from the MD runs you already have. But don't take structures which are
in the XDATCAR after each other. There should be some space between the extracted structures.
For the extracted structures you would have to make DFT calculations with the same settings
as when you trained the force field. Afterwards you would have to compute the root mean square
error between the DFT calculations and the MLFF calculations. You should compute the root mean square for the energy, the forces and the stress tensor.
You could also try to do a DFT MLFF comparison for the structures directly before your bond breaking occurs. In principle, if the error goes up the closer you get to the structure with broken bond the more certain it is that you have to explore more phase space during training.
Please let me know about your findings to discuss further.
All the best Jonathan