Trained LJ potential cannot reproduce DFT pressure

[lnnbig]

Hi Mingjian,

Thank you very much for your continued support and for answering my questions.

With the trained LJ coefficients for Si (sigma = 2.0629043239028659, epsilon = 1.5614870430532530) on dataset “Si_training_set_4_configs”, I performed a LAMMPS single-point calculation (on the first snapshot, i.e., Si_alat5.409_scale0.005_perturb1.xyz). I found that the LAMMPS pressure (-25 GPa) deviates significantly from my DFT reference pressure (3 GPa).

I understand that the large pressure prediction error may be related to that the dataset “Si_training_set_4_configs” does not have stress tensor information. So I additionally created a new dataset which explicitly includes stress tensors for LJ potential construction. But the pressure prediction accuracy from the trained LJ potential remains unsatisfactory.

Attached are the input & output files for my LAMMPS and VASP benchmark calculations. Could you please help me take a look to see if there are someting going very wrong?
lammps_lj_vs_vasp.zip

[mjwen]

Hi @lnnbig

I don't have the bandwidth to provide detailed answers. But here are some general rules:

• LJ is not a good model for Si. You may want to use other models, like Stillinger-Weber
• Si_training_set_4_configs is too small for training a useful model. In reality, you need hundreds to thousands data points to train a physics-based model

Hope it helps !