how to map structural descriptor to atomic energy

[J-You]

Hi all,

I have two questions that are confusing me, and I hope to find solutions:

1. To my understanding, the fitting neural network is a representation of PES from AIMD, and the point in PES should correspond to a frame. The total energy E of a frame is the sum of atomic energy Ei and we use fitting neural network to map structural descriptor D to get atomic energy Ei. The fitting neural network should be a supervised model, but how can we acquire the prediction target Ei? In training dataset, the energy is the total energy, and actually, Ei cannot be derived from VASP-AIMD simulations.
2. Is it possible to extract the output of embedding net, that is, the structural descriptor D from DPMD-kit?

thanks for any help,
Jiang

[njzjz]

the fitting neural network is a representation of PES from AIMD,

Actually no. Please read our paper. Eq. 1 and Eq.2 can be a representation of PES. The fitting neural network $\mathcal{F}$ is only a part of Eq. 1 and cannot be considered PES.

The fitting neural network should be a supervised model, but how can we acquire the prediction target Ei? In training dataset, the energy is the total energy, and actually, Ei cannot be derived from VASP-AIMD simulations.

I guess you mean how to fit $E_i$. Actually, the total energy is used in the loss function instead of $E_i$. See Eq. 60.

Is it possible to extract the output of embedding net, that is, the structural descriptor D from DPMD-kit?

See #1852.

[J-You]

thanks for your reply.

I guess you mean how to fit Ei Actually, the total energy is used in the loss function instead of Ei. See Eq. 60.

Sure, the predicted total energy is used in loss function, but I am confused how to fit Ei, because the true Ei is unknown, at least not in training dataset. The predicted total energy should be the sum of the predicted Ei? But I am not sure, it seems also an assumption for extensive character of DP.

[njzjz]

The decomposition of $E$ into $E_i$ is arbitrary and there can be countless different combinations (with the same total energies and forces). As a result, $E_i$ must have no physical meaning and thus cannot be fitted alone in theory.

[J-You]

Got it! I should look through your published paper. Thanks!