DFT energy and predicted energy off by 4eV

[MathildeAnd]

Hello everyone,

I'm doing ML from systems composed by Fe+H2O and FeO+H2 at 2000K, 3000K and 4000K and I trained the potential for each separate temperature. I encounter a problem when testing the model : the predicted energy and the DFT energy are off by approximately 4 eV. With my training data set composed by AIMD by VASP, I used LDA+U with a J parameter of 1 and a U of 5 so the gap in the energies seems to correspond to the LDA+U parameter. However I don't understand how to fix this problem, do you have any suggestions ?
I attached my input and the figure I obtain for the predicted energy against DFT energy

input_MyPotential.json

[njzjz]

I see that your training data contains systems with different number of atoms. Do you use the same LDA+U parameter for these systems? Do all of these systems have the same bias?

[njzjz]

Do you see the consistent error in lcurve.out?

[MathildeAnd]

Here is the lcurve.out file for 2000K system.

[njzjz]

It looks to me like some points have inconsistent 1e-4 rmse_e_trn, which is much smaller than the bias shown in the figure.

[njzjz]

Could you plot all the points in the training data instead of some of them?

[MathildeAnd]

You mean all of the 10.000 points ?
Could it be an overfitting problem ?

[wanghan-iapcm]

So you have trained three DP models, for 2000K, 3000K and 4000K, respectively. The labeling method is LDA+U with J=1 and U=5

What was the model for the test plot? How was the test dataset constructed?

[wanghan-iapcm]

how many atoms do you have in your configurations? could you please show the command line output of the dp test?

[MathildeAnd]

For Fe+H2O I have two different systems :

1. with 100 iron and 40 H2O so 220 atoms
2. with 100 iron and 70 H2O so 310 atoms
   For FeO+H2 I have also two systems :
3. with 256 FeO (half-half) and 50 H2 so 356 atoms
4. with 156 FeO (half-half) and 78 H2 so 312 atoms

I'm not sure what you mean by the command line output of the dp test but the final output command was this :
DEEPMD INFO Energy RMSE : 4.210666e+00 eV
DEEPMD INFO Energy RMSE/Natoms : 1.913939e-02 eV
DEEPMD INFO Force RMSE : 2.433533e-01 eV/A
DEEPMD INFO Virial RMSE : 6.228288e+00 eV
DEEPMD INFO Virial RMSE/Natoms : 2.831040e-02 eV

[wanghan-iapcm]

the std of your energy label per atom is approximately 8 / 300 = 26 meV/atom, which is relatively small.
the energy training error is 19 meV/atom, mainly contributed by the "constant shift"

to fix the issue it is recommended to refine your model with reduced force prefactors, e.g. 1-> 1 and increased energy prefactors e.g. 1->10 and a reduced start learning rate, e.g. 10 times smaller than from scratch training

you may use the init-model mode to load model parameters and start a refinement, see https://docs.deepmodeling.com/projects/deepmd/en/r2/train/training-advanced.html#options-and-environment-variables

[MathildeAnd]

Ok thank you for your suggestions, I tried to tweak this parameters but I still end up with a residual error on the forces of 1 eV...

[wanghan-iapcm]

if you find you cannot simultaneously improve the energy and force accuracies, you may check the quality of your data following https://docs.deepmodeling.com/projects/deepmd/en/r2/troubleshooting/precision.html