How to improve the the Dp model with low RMSE?

[Guancred]

I have built the dp model by the code of dpgen, then collecting the data for a long training (7000000), while the energy and force RMSE are 60 meV/atom and 350 meV/A.

So, if i want to improve this dp model and decrease the RMSE, which training parameters should be adjusted ?

Thank you very much!

[wanghan-iapcm]

you may try with dpa-1 (with attention layers) https://github.com/deepmodeling/deepmd-kit/blob/r2/examples/water/se_atten/input.json
and dpa-2 https://github.com/deepmodeling/deepmd-kit/blob/devel/examples/water/dpa2/input_torch.json

[Guancred]

Thank you, i have tried with the dpa-1, while the RMSE still show a large value,

and the input.json:
{
"model": {
"descriptor": {
"type": "se_atten",
"sel": 140,
"rcut_smth": 0.5,
"rcut": 7,
"neuron": [
25,
50,
100
],
"resnet_dt": false,
"axis_neuron": 16,
"attn":128,
"attn_layer":2,
"attn_mask":false,
"attn_dotr":true,
"type_one_side": true,
"seed": 1541804686,
"_activation_function": "tanh"
},
"fitting_net": {
"neuron": [
240,
240,
240
],
"resnet_dt": true,
"seed": 1477278577

Any parameters should be revised for the decreased RMSE? Thanks very much!

[wanghan-iapcm]

apart from the mentioned model architectures, you may follow the doc https://docs.deepmodeling.com/projects/deepmd/en/r2/troubleshooting/precision.html#why-does-a-model-have-low-precision to improve the accuracy of the model

[Guancred]

Hi, whether the "compress" is not support for the DPA-1 model? when i compress it, an error will be occur: "can not compress model when attention layer is not 0."

[wanghan-iapcm]

only compressible when attn_layer is set to 0

[Guancred]

Thanks, so, i should set the "attn_layer: 0", and then train another model if i want to compress?