How to normalize the descriptors for deepmd?

[lhycms]

It seem that we should use davg and dstd to normalize descriptors before training, and the inference process also requires these two data. But i can't figure out how to calculate davg and dstd.
Are they the statistic data for (s, sx/r, sy/r, sz/r)? But the feature of deepmd is G^T R R^T G, if we just normalize (s, sx/r, sy/r, s*z/r), we are unable to guarantee features normalized.
Looking forward to everyone joining the discussion!

[njzjz]

davg and dstd is only used to normalize the environmental matrix $\mathcal{R}^i$. $\mathcal{G}^i$ is derived from $\mathcal{R}^i$, so we expect it does not need to be normalized. $\mathcal{D}^i$ is given by

$$\mathcal{D}^i = \frac{1}{N_c^2} (\mathcal{G}^i)^T \mathcal{R}^i (\mathcal{R}^i)^T \mathcal{G}^i_<$$

[lhycms]

Thanks for th help!