Systematic energy offsets between NN and DFT results

[LiangMD-BGI]

Hello all,

While forces are consistent between results from NN potential and DFT,
there is a systematic offset of energy between NN and DFT results.

Below is my input.json file for deepmd:

        "type": "se_a",
        "sel": [
            100,
            100
        ],
        "rcut_smth": 0.5,
        "rcut": 10.0,
        "neuron": [
            25,
            50,
            100
        ],
        "axis_neuron": 12,
        "seed": 3246556277
    },
    "fitting_net": {
        "neuron": [
            100,
            100,
            100
        ],
        "resnet_dt": true,
        "seed": 1978363284
    }
},
"learning_rate": {
    "start_lr": 0.001,
    "decay_steps": 5000,
    "decay_rate": 0.95,
    "_comment": " that's all"
},
"loss": {
    "start_pref_e": 1,
    "limit_pref_e": 1,
    "start_pref_f": 1000,
    "limit_pref_f": 100,
    "start_pref_v": 1,
    "limit_pref_v": 1
},
"training": {
    "systems": [
        "../data/"
    ],
    "set_prefix": "set",
    "stop_batch": 20000000,
    "batch_size": [
        1
    ],
    "disp_file": "lcurve.out",
    "disp_freq": 100,
    "numb_test": 4,
    "save_freq": 1000,
    "save_ckpt": "model.ckpt",
    "load_ckpt": "model.ckpt",
    "disp_training": true,
    "time_training": true,
    "profiling": false,
    "profiling_file": "timeline.json",
    "seed": 2662107775
}

}

[LiangMD-BGI]

more information on lcurve file:

batch l2_tst l2_trn l2_e_tst l2_e_trn l2_f_tst l2_f_trn l2_v_tst l2_v_trn lr
first line: 0 1.34e+02 1.31e+02 8.30e+00 8.62e+00 1.64e+00 1.04e+00 1.49e+00 2.19e-01 1.0e-03
latest line: 1153400 4.05e-01 4.51e-01 9.16e-03 9.63e-03 1.52e-02 1.80e-02 8.32e-03 8.53e-04 5.2e-04

[finthon]

I also found the same problem, there is a systematic offset of energy between NN and DFT results, How did you fix this issue finally? @LiangMD-BGI

[LiangMD-BGI]

Hi, I have not figured out this problem. I also found that, during training, if you freeze the model at different times, the predicted NN energy could be either systematically larger or smaller than the one from DFT. So I guess this error is related to the internal error of NN fitting. Looking forwards to having more discussion with you.

[njzjz]

I notice you set both start_pref_e and limit_pref_e to 1. You may consider to increase them although it may decrease the accuracy of forces. (Here energies and forces are inconsistent)

[LiangMD-BGI]

I notice you set both start_pref_e and limit_pref_e to 1. You may consider to increase them although it may decrease the accuracy of forces. (Here energies and forces are inconsistent)

Hi Jinzhe, thanks for your suggestions. I will try it and see whether the results get better.

[Hashan-Peiris]

I am facing the same issue. Was it fixed by any chance?

[J-You]

I want to know if it is suitable to set "rcut_smooth" as 0.5 in your input.json file?
Maybe it is equal to 9.5 according to your "rcut" setting?

[LiangMD-BGI]

The rcut_smth gives where the smoothing starts. The setting (rcut_smth = 0.5 and rcut = 10), therefore, allows smoothing in a larger range of radius relative to the one you suggested (rcut_smth = 9.5 and rcut = 10).
Such a small rcut_smth = 0.5 is present in several works (e.g., https://www.pnas.org/content/117/42/26040.short).
So there should be no problems there.

[J-You]

Thanks for your response. It extends my understanding scope.

[mirza-galib]

Hi, I am not sure if the offset between DFT energy and NN energy problem is solved. You are getting this because your weight factor on the force is high and on the energy is low. You need to end with equal weight on force and energy to improve the accuracy on energy training.

[LiangMD-BGI]

Hi, thank you very much for your suggestion. I will try it and see whether the results get better.

[sooyaaa233]

Did you solve your problem now? How did you solve it?