Force RMSE in multi component system

[ald3hyde]

I am currently trying to create a multi-component model of a 5-element system. I have encountered a strange phenomenon: the RMSE behavior for the 3-component system is natural for the supervised data, but the RMSE behavior for the 4- and 5-component systems is off by a certain rule. Could this be caused by AIMD result ?

All data are at high temperatures (3000K).
The data for the plots were generated by dp test.
・5 component system

・3 component system

・my input.json script
"descriptor": {
"type": "se_atten",
"sel": 110,
"rcut": 6,
"rcut_smth": 0.5,
"neuron": [
25,
50,
100
],
"axis_neuron": 16,
"resnet_dt": false,
"seed": 0,
"attn": 120,
"attn_layer": 0,
"attn_dotr": true,
"attn_mask": false
},
"fitting_net": {
"type": "ener",
"neuron": [
240,
240,
240
],
"resnet_dt": true,
"seed": 1
}
},
"learning_rate": {
"type": "exp",
"start_lr": 0.001,
"stop_lr": 3.51e-8,
"decay_steps": 5000
},
"loss": {
"type": "ener",
"start_pref_e": 0.02,
"limit_pref_e": 1,
"start_pref_f": 10000,
"limit_pref_f": 1,
"start_pref_v": 0,
"limit_pref_v": 0,
"relative_f": 10

[njzjz]

Could you show a sample of the data?

[ald3hyde]

show you maximum force values at each step
・5-component(2000K)(Extracted every 10steps after 5000 step)

・3-component(3000K)(Extracted every 10steps after 2000 step)

[njzjz]

These figures do not contain much information. It will be helpful if you show a structure with a strange error. I guess there is interaction in the range of more than 6 Å in these structures.

[ald3hyde]

The target is in a molten salt state. Strangely enough, the structure (RDF) is very good, but when fitting to multi-component (with sodium) data, it ends up being the same. However, when fitting multi-component (with sodium) data, I get the same results. Is this a result of long-range interaction due to sodium?