questions about DeepMD permutation symmetry & network

[chandagsin]

Dear Developers,

one of the key points of DeepMD is to construct a descriptor that preserves the permutation symmetry. I am new to neural network. It appears to me that you have proposed two different descriptors for the permutation symmetry (and I am a bit confused about this).

(1) {D_ij} = {1/r_ij, x_ij/r_ij, y_ij/r_ij, z_ij/r_ij}, (e.g, Wang et al, DeepMD-kit, Comp. Physics Comm. 228 (2018) 178–184)
(2) a network that maps the original local coordinate to the descriptor. the descriptor satisfies the permutation symmetry by its own structure. (e.g., Zhang et al., 2018 DeepPot-SE )

From what I understand, (1) itself is not permutation symmetric unless enough samples are taken from the configuration space, since molecules are not perfectly rigid, e.g., a CH4 molecule may not be a regular tetrahedron (may be distorted and hence the H- atoms symmetry is broken) at a specific point of time.

I would like to know, if the current DeepMD-kit actually adopts (2). the example JSON files specify the descriptor as (e.g.),
"descriptor" :{
"type": "se_e2_a",
"sel": [46, 92],
"rcut_smth": 5.80,
"rcut": 6.00,
"neuron": [25, 50, 100],
"resnet_dt": false,
"axis_neuron": 16,
"seed": 1,
"_comment": " that's all"
},

I think this "neuron" would be the network for the construction of descriptor, as detailed in Zhang et al., 2018.

I also want to know what would be proper sizes ("neuron") for the embedding net and the fitting net?

I guess the "axis_neuron" has something to do with the "G" matrix in Zhang's paper. what would be a good value for it?

Thanks!

[njzjz]

Both 1 (loc_frame) and 2 (se_e2_a) are available, and the descriptor is decided by setting descriptor/type. For more information, please read documentation.

For neuron please read the FAQ.

[chandagsin]

Thanks!. for Axis_neuron, I guess it's the number of columns of matrix G. what would be a good value?

[njzjz]

We've never benchmarked axis_neuron. You can start with the value that you posted.