Page 1 of 1
ML_REG format
Posted: Fri Oct 21, 2022 7:21 am
by duccoi
Hello, I found in the VASP wiki: "This file contains the regression results of the force field training compared to the ab-initio training data." Can anyone please help me confirm the format of the ML_REG file:
The energy part of two columns: Column1- Energy F obtained by MLFF; and Column 2 - the energy obtained by ab into.
The force part consists of two columns: Column 1 - Average force obtained by MLFF and Column 2 - the average force obtained by Ab initio.
One more question. I trained MLFF with 20,000 steps, then I found about 200 lines were printed out for the energy part, which means only 200 steps included in the comparison between MLFF and Ab Initio. Some may comment that more comparisons should be done. Which parameter should I control to increase more lines (maybe 1000-2000 lines) written in energy parts of ML_REG (I tried to increase NSW, but it slightly increase)?
Re: ML_REG format
Posted: Tue Oct 25, 2022 10:01 am
by ferenc_karsai
We have updated the ML_REG page on the VASP wiki:
wiki/index.php/ML_REG
The first column is always the ab-initio data. The second is the fitted data. The forces are not averaged forces, but all force components for eacht training structure listed after each other.
The number of entries for the energies corresponds to the number of training structures. That means "The number of configurations" in the corresponding ML_ABN file. This should be all "grep STATUS ML_LOGFILE" lines that contain 'threshold, learning and critical'.
In the on-the-fly learning this is much less than the number of MD steps set by NSW.
Re: ML_REG format
Posted: Tue Nov 01, 2022 2:10 am
by duccoi
Dear @ferenc_karsai,
Thank you for your clarification.
For the number of configurations for the training process, is it impossible to control the number of training structures with parameters from INCAR such as NSW or ML_MCONF?
Re: ML_REG format
Posted: Thu Nov 03, 2022 8:11 am
by ferenc_karsai
Actually there is no way to say in advance how many training structures one will have, since it is determined by the on-the-fly learning. Hence there is also no way of controlling the training structures directly. Rather you can influence the frequency of the sampling of local reference structures and hence also training structures (since training structures are only sampled if at least one new local reference configuration is taken) as described here
wiki/index.php/Best_practices_for_machi ... parameters
Re: ML_REG format
Posted: Sun Nov 06, 2022 1:50 pm
by duccoi
Dear @ferenc_karsai,
Thank you so much for your answer. I tried to train the force field with a structure of 140 atoms (an oxide of transition metals). I plotted data written in ML_REG files and found that the differences between energies obtained by Ab initio and MLFF are still large even though I trained several times. Could you please suggest to me to choose suitable parameters in the INCAR file?
Code: Select all
ISTART = 1
LWAVE = .FALSE.
LCHARG = .FALSE.
NCORE = 8
LREAL= .F.
ICHARG = 1
LORBIT = 10
KSPACING = 0.5
ENCUT = 450 eV
ISPIN = 2
LMAXMIX= 4
MAGMOM = 3*0 1 15*0 1 16*0 72*0 36*0
GGA = PE
ISMEAR = 0
SIGMA = 0.02
IALGO = 48
NELMIN = 3
NELM = 300
EDIFF = 1E-07
ML_LMLFF = .TRUE.
ML_ISTART = 1
RANDOM_SEED = 688344966 0 0
ML_MCONF = 10000
ML_IWEIGHT = 2
ML_WTOTEN = 2
ML_IALGO_LINREG = 1
ML_SION1 = 0.3
ML_MRB1 = 12
ISIF = 2
TEBEG = 400
TEEND = 400
IBRION = 0
ISYM = 0
NSW = 50000
NBLOCK = 1
MDALGO = 2
SMASS = 1.0
POTIM = 2.0
IALGO = 48
Code: Select all
# ERR ######################################################################
# ERR This line contains the RMSEs of the predictions with respect to ab initio results for the training data.
# ERR
# ERR nstep ......... MD time step or input structure counter
# ERR rmse_energy ... RMSE of energies (eV atom^-1)
# ERR rmse_force .... RMSE of forces (eV Angst^-1)
# ERR rmse_stress ... RMSE of stress (kB)
# ERR ######################################################################
# ERR nstep rmse_energy rmse_force rmse_stress
# ERR 2 3 4 5
# ERR ######################################################################
ERR 0 4.50398330E-03 3.58666632E-02 8.56463489E-01
ERR 1 4.50504300E-03 3.56574344E-02 8.53957528E-01
ERR 2 4.50567464E-03 3.56811046E-02 8.54435940E-01
ERR 3 4.50624798E-03 3.56109230E-02 8.53375338E-01
ERR 4 4.50681918E-03 3.55859759E-02 8.52932644E-01
ERR 5 4.50739952E-03 3.55715772E-02 8.52425643E-01
ERR 6 4.50785073E-03 3.55533846E-02 8.51737706E-01
ERR 7 4.50834818E-03 3.55499122E-02 8.51893755E-01
ERR 8 4.50880821E-03 3.55221450E-02 8.51464303E-01
ERR 9 4.50938291E-03 3.54937760E-02 8.51365065E-01
....
ERR 16591 4.55124994E-03 3.48468351E-02 8.41362815E-01
ERR 17294 4.55219874E-03 3.48278895E-02 8.41119079E-01
ERR 19507 4.55310823E-03 3.48100958E-02 8.40667789E-01
ERR 22622 4.55413019E-03 3.47888100E-02 8.40398921E-01
ERR 25720 4.55511057E-03 3.47763432E-02 8.39986837E-01
ERR 30414 4.55566064E-03 3.47635375E-02 8.40297753E-01
ERR 32528 4.55626462E-03 3.47542045E-02 8.39833238E-01
ERR 34126 4.55662674E-03 3.47410917E-02 8.39368749E-01
ERR 42067 4.55741335E-03 3.47222220E-02 8.39830052E-01
ERR 42132 4.55770161E-03 3.47102062E-02 8.39746328E-01
ERR 50000 4.55773958E-03 3.47022296E-02 8.39277708E-01
Code: Select all
# BEEF ########################################################################################################################
# BEEF This line shows the Bayesian error estimations and the current threshold criterion.
# BEEF
# BEEF nstep ............ MD time step or input structure counter
# BEEF bee_energy ....... BEE of energy per atom (eV atom^-1)
# BEEF bee_max_force .... BEE of forces (max) (eV Angst^-1)
# BEEF bee_ave_force .... BEE of forces (average) (kB)
# BEEF threshold ........ Current value of threshold criterion (eV Angst^-1)
# BEEF bee_max_stress ... BEE of stresses (max) (kB)
# BEEF bee_ave_stress ... BEE of stresses (average) (kB)
# BEEF ########################################################################################################################
# BEEF nstep bee_energy bee_max_force bee_ave_force threshold bee_max_stress bee_ave_stress
# BEEF 2 3 4 5 6 7 8
# BEEF ########################################################################################################################
BEEF 0 0.00000000E+00 0.00000000E+00 0.00000000E+00 2.00000000E-03 0.00000000E+00 0.00000000E+00
BEEF 1 1.39735472E-07 1.07822273E-02 2.86131956E-03 2.00000000E-03 8.25802901E-02 6.26389645E-02
BEEF 2 1.36786793E-07 1.06018412E-02 2.83074079E-03 2.00000000E-03 8.12703047E-02 6.12463379E-02
BEEF 3 1.34562019E-07 1.03259493E-02 2.81156988E-03 2.00000000E-03 8.06198930E-02 6.05198244E-02
BEEF 4 1.32202624E-07 9.94784054E-03 2.78776478E-03 2.00000000E-03 8.05619193E-02 6.02620460E-02
BEEF 5 1.29910878E-07 9.78424950E-03 2.76091370E-03 2.00000000E-03 8.11237187E-02 6.04451403E-02
BEEF 6 1.27612149E-07 9.60263153E-03 2.72981541E-03 2.00000000E-03 8.21944973E-02 6.09242134E-02
BEEF 7 1.25224973E-07 9.06388980E-03 2.69457926E-03 2.00000000E-03 8.35069183E-02 6.14642393E-02
BEEF 8 1.23300230E-07 8.25745732E-03 2.66165857E-03 2.00000000E-03 8.45500437E-02 6.18008615E-02
BEEF 9 1.22073272E-07 7.57083201E-03 2.63657115E-03 2.00000000E-03 8.49832480E-02 6.17298302E-02
BEEF 10 1.22106800E-07 7.57999916E-03 2.62491346E-03 9.35169177E-03 8.46277889E-02 6.12287029E-02
BEEF 11 1.22879121E-07 7.77161446E-03 2.62547172E-03 9.35169177E-03 8.44858031E-02 6.09907472E-02
BEEF 12 1.23515590E-07 8.08440552E-03 2.63659549E-03 9.35169177E-03 8.37660690E-02 6.06099015E-02
BEEF 13 1.23910179E-07 8.44539519E-03 2.65555459E-03 9.35169177E-03 8.26577059E-02 6.02257469E-02
BEEF 14 1.24430393E-07 8.78036399E-03 2.67965326E-03 9.35169177E-03 8.14788303E-02 5.99549887E-02
BEEF 15 1.25259198E-07 9.21077155E-03 2.70577228E-03 9.35169177E-03 8.05633248E-02 5.98614648E-02
BEEF 16 1.26462613E-07 9.74337988E-03 2.73052150E-03 9.35169177E-03 8.01507195E-02 5.99602311E-02
....
BEEF 49989 1.46464070E-07 1.12690722E-02 3.10842446E-03 2.36818029E-02 8.90924362E-02 6.83355420E-02
BEEF 49990 1.46485741E-07 1.13239641E-02 3.10832900E-03 2.36818029E-02 9.12662501E-02 6.95469181E-02
BEEF 49991 1.45821690E-07 1.09594351E-02 3.10200998E-03 2.36818029E-02 9.29271753E-02 7.04206216E-02
BEEF 49992 1.44683433E-07 1.02993363E-02 3.09328887E-03 2.36818029E-02 9.39407177E-02 7.08848537E-02
BEEF 49993 1.43626199E-07 1.06683781E-02 3.08638738E-03 2.36818029E-02 9.44075741E-02 7.10447587E-02
BEEF 49994 1.42973700E-07 1.09000059E-02 3.08409728E-03 2.36818029E-02 9.43294030E-02 7.09667569E-02
BEEF 49995 1.43060059E-07 1.09784856E-02 3.08879644E-03 2.36818029E-02 9.37458122E-02 7.07321439E-02
BEEF 49996 1.43880941E-07 1.09132546E-02 3.10332632E-03 2.36818029E-02 9.30897857E-02 7.05804205E-02
BEEF 49997 1.45639880E-07 1.07315605E-02 3.12837659E-03 2.36818029E-02 9.30941661E-02 7.08392422E-02
BEEF 49998 1.48160513E-07 1.08187197E-02 3.15925015E-03 2.36818029E-02 9.41416464E-02 7.16372492E-02
BEEF 49999 1.50508802E-07 1.12654710E-02 3.18699796E-03 2.36818029E-02 9.58798703E-02 7.27362433E-02
BEEF 50000 1.52001293E-07 1.15908456E-02 3.20189940E-03 2.29205530E-02 9.73523558E-02 7.35974466E-02
Re: ML_REG format
Posted: Mon Nov 07, 2022 10:34 am
by ferenc_karsai
Your root means square error of 4.5 meV/atom is not bad. Especially the error of 30 meV/Angstrom for the forces is great.
But it also depends on how good the force field is on an independent test set.
We have a page for best practices in machine learning together with molecular dynamics:
https://www.vasp.at/wiki/index.php/Best ... rce_fields
Basically you are not employing many of the advises on that page. Nevertheless your force-field is quite good. I suggest you study that page and employ the things on that page to further improve your force field.
Here are some key points you should definitely try out:
-) ML_WTOTEN: Weight energy stronger (factor 10-50). This will worsen the forces and stress fits a bit, but can really greatly increase the accuracy on the energy. So this trade-off is worth it.
-) Accurate force fields: Complete on the fly training with ML_IALGO_LINREG=1; ML_SION1=0.3; ML_MRB1=12 and then refit using NSW=0; ML_IALGO_LINREG=3; ML_CTIFOR=1000.0; ML_EPS_LOW=1.0E-14; ML_SION1=0.5; ML_MRB1= 8.
-) To get a more stable force field use run in an NpT ensemble (MDALGO=3, ISIF=3 and remaining Langevin parameters) and definitely use a temperature ramp. You wrote that it only picked out 200 training structures (and possibly also not many local reference configurations). The temperature ramp and fluctuating lattice should increase the numbers.
-) Possibly try to improve the ab-initio data, since the more accurate the ab-initio data the easier it is to fit it. Increase ENCUT to let's say 550 eV (needs to be anyway larger for an NpT ensemble). You haven't written how large your supercell is but if it is not huge then possibly increase the k-spacing, since you set KSPACING = 0.5 and that results in a very low number of k points.
I hope you will see some improvement.