How does a frozen machine learning potential accurately identify element types in LAMMPS simulations?

[jzh-w]

When using machine learning potentials for LAMMPS simulations, regardless of whether pair_coeff is set as pair_coeff * * or pair_coeff * * B B O O, one element type is recognized as another element type during the simulation. How should I configure this to avoid such issues?

[jzh-w]

The first image corresponds to pair_coeff * * B B O O, and the second image corresponds to pair_coeff * *. This situation occurred after relaxation at 300K for 10 ps. I suspect that in the first case, O was identified as B, while in the second case, the B and O in boron oxide were misidentified, resulting in this situation during the relaxation process.

[jzh-w]

I'm very sorry, I misspoke. The first image corresponds to pair_coeff * *, and the second image corresponds to pair_coeff * * B B O O.