Unusually High RMSEs Running ML on Bulk Water

[ramoser]

Hello,

I am having problems running machine learning on a bulk water system. I am using the NpT ensemble along with an ICONST file (provided below) to have a system with a variable volume but a fixed shape. The system runs as I expect, but the RMSEs for the energy are about 600 meV/atom when I expect values of about 2-10 meV/atom. I also previously tried NVT to no avail. Is there an effective way to lower the energy RMSEs without drastically changing the system?

[ahampel]

Hi,

thank you for reaching out on the official VASP forum. Sorry for the late reply but I tried over the weekend to analyze your calculation and tweak it to better understand what is going on.

First of all an error of 600 meV/atom is of course much too high an the ML force field is not good here. The only thing in the INCAR catching my eye was IDIPOL=3? This will include dipole corrections along z-axis in the total energy. But I am not sure what the relevance is for bulk systems? Maybe I missed something? And also I am not sure if it is necessary to change the unit cell during MD with ISIF=3 maybe ISIF=2 is sufficient (or ISIF=8 to change volume but not shape of cell).

During the MD I noticed that the RMSE values are first very good but then jump to higher values:

Code: Select all

ERR                     1   5.53863475E-03   5.79626930E-02   4.39148731E+01
ERR                     2   1.07326182E-02   6.20710905E-02   1.82238709E-01
ERR                     3   1.01000572E-02   5.54394459E-02   5.73041546E-01
ERR                     4   9.09089039E-03   5.65879264E-02   8.39240041E-01
ERR                     5   8.29047026E-03   6.07126961E-02   9.26056392E-01
ERR                     6   7.63009874E-03   5.48429810E-02   1.17942402E+00
ERR                     7   7.09365402E-03   5.24972488E-02   1.23637912E+00
ERR                     8   6.67823073E-03   5.42941938E-02   1.27551659E+00
ERR                     9   6.32694996E-03   5.10759618E-02   1.27680210E+00
ERR                    10   6.01139440E-03   5.08182276E-02   1.25051736E+00
ERR                    20   5.79190330E-03   5.44587922E-02   1.11429015E+00
ERR                    70   1.99989519E-02   4.25879285E-02   5.64779008E-01
ERR                   120   6.62002670E-01   4.36455072E-02   7.29601084E-01
ERR                   155   6.09453843E-01   4.68924683E-02   8.83649052E-01
ERR                   205   5.57664183E-01   4.90052227E-02   1.06580800E+00
...

So the first 50 steps of the MD seem to be okay. And if I look at the temperature of the system in each MD step it seems that after 50 steps or so the temperature suddenly goes up and is not around 313 K. So maybe the MD itself has problems here, and this might be not related to how the ML is setup, i.e. the system does not settles at this temperature and something odd is happening.

One thing I now tried is to artificially increase the mass of the H atoms: OMASS = 16.0 6.0 from 1.0 to 6.0 and remove the IDIPOL=3 line. I also lowered a bit the energy threshold for the SCF steps: EDIFF = 1.0E-7 to increase accuracy of the forces. This seems to help:

Code: Select all

ERR                     1   4.98995150E-03   6.10057736E-02   6.41476315E+01
ERR                     2   7.01219129E-05   1.05594347E-01   3.61017304E-01
ERR                     3   1.01207332E-04   1.15176487E-01   4.46667249E-01
ERR                     4   1.07391994E-04   1.19369121E-01   3.83945851E-01
ERR                     5   9.76030786E-05   1.11387401E-01   2.72005220E-01
ERR                     6   8.92927797E-05   1.04646508E-01   3.96284033E-01
ERR                     7   7.63095799E-05   9.55711465E-02   6.06757848E-01
ERR                     8   6.71448288E-05   8.81478044E-02   7.57363918E-01
ERR                     9   5.88291799E-05   8.31446566E-02   8.52951330E-01
ERR                    10   5.34249788E-05   7.71431783E-02   9.08014456E-01
ERR                    51   1.19782193E-04   6.25488505E-02   1.02527036E+00
ERR                   101   1.19662148E-04   4.69869070E-02   1.16926833E+00
ERR                   168   1.27986932E-04   5.18966058E-02   1.25195540E+00
ERR                   220   2.07319369E-04   5.45972466E-02   1.22579630E+00
ERR                   273   2.31123738E-04   5.54158360E-02   1.01971706E+00
ERR                   331   2.11360340E-04   5.82848683E-02   1.00644982E+00
ERR                   381   1.81875291E-04   5.95726926E-02   9.26518766E-01
ERR                   431   1.80875725E-04   6.10994395E-02   8.98983131E-01
ERR                   481   1.69580993E-04   6.24835734E-02   8.27475970E-01
ERR                   531   1.77779285E-04   6.34564061E-02   8.22247270E-01
ERR                   754   2.33512177E-04   6.51511773E-02   8.23173184E-01
ERR                   840   2.60944852E-04   6.48715992E-02   8.60442859E-01
ERR                  1012   2.45628800E-04   6.55058552E-02   8.50335030E-01
ERR                  1207   2.75715888E-04   6.58006168E-02   8.02750955E-01
ERR                  1607   3.54000582E-04   6.53093885E-02   7.93547860E-01
ERR                  1663   3.66643802E-04   6.50961097E-02   8.06753652E-01
ERR                  1752   3.95937513E-04   6.52459855E-02   7.76386093E-01

but the temperature is much higher now around 370 K at later MD steps. Maybe you can try this and also check if the MD itself behaves as expected.

Best,
Alex

[ahampel]

Hi,

is your problem fixed? Let me know if you have any further problems. Otherwise I will close this issue soon.

Best regards,
Alex