generate a ML-FF for CO2 molecules

[wuyang_lin]

Dear VASP developers, 

 

We are trying to generate a new machine learning force field for CO2 adsorption on copper surfaces. We have finished the training for bulk Cu and several surfaces and the results compare well with ab initio MD and DFT optimizations. However, once we take CO2 molecules into account, problems arise: The ML-FF results show that an isolated CO2 molecules can not keep linear and the angle within the molecule is always between 165° and 180°. Besides, we tried to add a constraint condition in ICONST file to fix the angle to 180°, which will cause calculation fail at once and RATTLE algorithm convergence isn’t achieved. Last, if we use a DFT-optimized configuration with CO2 adsorbed on copper surface to train the force field, the results are also inaccurate and will make the force field less accurate for pure copper. 

 

We have also tried to simulate liquid CO2 but we again observe bending of molecules. 

 

Could you please help us find out what could be the reasons for getting bent CO2?

Many thanks,
Wuyang

[alexey.tal]

Dear Wuyang,

Could you please provide all the relevant input and output files for your calculation according to the forum guidelines.

[wuyang_lin]

Dear Alexey,

Thank you very much for you reply. Please find attached the relevant files.

Wuyang

[alexey.tal]

In the provided files, the temperature grows from 300 K up to 1800 K, which indicates that the thermostat was not set correctly.
Indeed, in the INCAR files, the parameters for a Langevin thermostat are provided, however, MDALGO = 2 is set, which corresponds to the Nose-Hoover thermostat.
If you would like to perform an NVT simulation with the Nose-Hoover thermostat, you should set the SMASS tag.
Please see the wiki MDALGO and SMASS