Large force RMSE

[melicherova]

Hi,
I have DFT MD trajectories from melting crystal nitrogen and liquid cooling under pressure
and I want to train NN with the data from this simulation.

But I have problem with large force RMSE in trainning which is 913 meV/A (energy RMSE is quite reasonable 9 meV/atom).
I used trained NN in lammps MD and I tried to reproduce protocol from DFT MD but crystal melted at much lower temperature
then in DFT MD.

Are there some options how to reduce force RMSE?

Thanks

I am using deepmd-kit v2.0.3 and input for deepmd is:

"model": {
    "type_map":     ["N"],
    "descriptor" :{
        "type":             "se_e2_a",
        "sel":              [190],
        "rcut_smth":   5.80,
        "rcut":             6.00,
        "neuron":           [25, 50, 100],
        "resnet_dt":        false,
        "axis_neuron":      16,
        "seed":             1,
        "_comment":         " that's all"
    },
    "fitting_net" : {
        "neuron":           [240, 240, 240],
        "resnet_dt":        true,
        "seed":             1,
        "_comment":         " that's all"
    },
    "_comment":     " that's all"
},

"learning_rate" :{
    "type":         "exp",
    "decay_steps":  5000,
    "start_lr":     0.001,
    "stop_lr":      3.51e-8,
    "_comment":     "that's all"
},

"loss" :{
    "type":         "ener",
    "start_pref_e": 0.02,
    "limit_pref_e": 1,
    "start_pref_f": 1000,
    "limit_pref_f": 1,
    "start_pref_v": 0.2,
    "limit_pref_v": 1,
    "_comment":     " that's all"
},

[McAlexDudu]

I also have this kind of question. My force RMSE is 55 meV/A which should at most 5 meV/A. I use dpgen can solve the problem; but I also want to know how to decrease the RMSE while using just deepmd. I have tried larger sel but not work.

By the way, my rmse_f_val from lcurve.out seem to almost keep in 40~60 meV/A from after 10000 steps, and seem to not converged?

[RobinMAJ]

Dear McAlexDudu
Could you provide the way on how dpgen solves this problem? Thank you very much!!

[McAlexDudu]

@RobinMAJ I think my problem maybe the dataset not set right, or my AIMD data were not good. During dpgen, I follow the step of init_bulk to create dataset, and the RMSE can decrease to at least 4 meV/A of four models.

[njzjz]

Could you provide the scatter plot (DFT vs NN)? I suggest that training data is extremely wrong.

[wanghan-iapcm]

Yes that could be a issue. One may want to check the shape of the simulation cell.

Could you also please provide the learning curves? Please show them in a log-log plot. Thank you!

[melicherova]

Yes, there is the learning curves plot.

[wanghan-iapcm]

The decaying speed of energy and virial looks perfect.

Could you also please check the RMAX parameter in your POTCAR and the minimal distance between atoms in the training data? Since your training data is at high pressure, I suspect there is an issue of core overlapping. http://jwz.chem.dal.ca/lab-notes.php

[melicherova]

RMAX is 1.529 Bohr and minimal distance between atoms is 1.0008 A in the training data.
However, this short distance comes from very small amount of N2 molecules in liquid structures.
I made two testing calculations with hard pseudopotential with RMAX=1.127. First one was with amorphous nitrogen and there was force deviation 334 meV/A compared to original calculation. Second one was with crystal and force deviation was 272 meV/A. Yes, it is quite large deviation but it does not explain difference in force RMSE in training where force RMSE for crystalline structures is 340 meV/A and for liquid and amorphous structures is ~900 meV/A.

[wanghan-iapcm]

The force deviation about ~300 meV/A is significant. The DP model assumes that the force is a derivative of the energy. The error introduced by core-overlapping, however, does not preserve the correlation. By far I cannot tell if all the deviation ~300 meV should be attributed to the error. In the worse case, the 300meV is a random error in the force label, it may draw DP from the correct force by 300 meV, and adding the 300meV random error at the test point, this could lead to at most ~600meV error in test. One meaningful test is to exclude all data frames that have short atomic distances, and train the model again to see the error in force.

From the point of using PP in the vasp calculation, since you have N2 molecules in the data, it is strongly recommended to use the harder PP with RMAC=1.127 Bohr. Of course the convergence with respect to ENCUT should be carefully tested. Harder PP needs larger ENCUT.

[RobinMAJ]

Thank you very much! 在2022-03-20 12:46:26，Han ***@***.***写道： The standard deviation of DFT forces look very large (~20 eV/A ?). When we talk about the accuracy of the model, we compare the error with respect to the standard deviation in the label. In your case the force error is about 1/10 of the standard deviation, your model has accuracy. — Reply to this email directly, view it on GitHub, or unsubscribe. Triage notifications on the go with GitHub Mobile for iOS or Android. You are receiving this because you were mentioned.Message ID: ***@***.***>