Train model on custom data?

[PythonFZ]

Hey PyXtal_FF team,

Is there any documentation available on how to create the input data for PyXtal_FF? The documentation starts with the provided JSON files but unfortunately I'm not familiar with the format or how to generate it.

I do have a list of ASE Atoms objects, with energies, forces and stresses available. The structure has periodic boundary conditions. What would be the best way to convert this data into a suitable PyXtal_FF input file?

Best Fabian

[qzhu2017]

@PythonFZ Hi, ASE is actually the most convenient way for us. So, if you have ase.db file, just created some additional tag as follows,

    data = {'dft_energy': eng, #total energy
            'dft_force': forces,     #N*3 array
            'dft_stress': stress,    #1*6 stress [in GPa, xx, yy, zz, xy, xz, yz]
            'group': group}

    with connect(db_name) as db:
        db.write(struc, data=data)

Please note that ASE and VASP output the stress tensor in different orders and units. Make sure you follow the order of

xx, yy, zz, xy, xz, yz

and the unit of Gpa

The following code should give you some idea to organize you ase data https://github.com/qzhu2017/PyXtal_FF/blob/32cc4e7598a00a04c304e6ba15151a136d3177ca/pyxtal_ff/utilities/__init__.py#L596-L636

[qzhu2017]

• [ ] to update the manual regarding the preparation of training data

[Gaurav-Arora-MatSci]

Hello @PythonFZ Can you guide me in generating manual input data for training the potentials? I am not very familiar with ase or any other package. It would be very helpful if you could guide me through the beginning process.

Thank you

[PythonFZ]

Hello @PythonFZ Can you guide me in generating manual input data for training the potentials? I am not very familiar with ase or any other package. It would be very helpful if you could guide me through the beginning process.

Thank you

I'm not using PyXtal Force Fields so I can't help you with this specific package.

[Gaurav-Arora-MatSci]

Hello @PythonFZ Just sent you an email.