How does MACE deals with DFT calculations with pressure? #1343
Replies: 18 comments
|
It really shouldn't care, unless your reference data is some sort of "net" stress, adding the internal virial stress and the applied stress, or enthalpy instead of potential energy. Assuming you're just converting from OUTCARs to xyz's with ASE I don't think that should happen, but you should check. |
|
Yeah it looks like you might be using enthalpy (energy+PV) instead of just free energy (electronic free energy) in your reference values. In the stresses you might be including the kinetic term in your reference, again you shouldn't |
|
Thanks for the quick answer @bernstei . I'm converting it from the vasprun.xml file and using energy for the REF_energy, not free_energy. |
|
Can you see if your "energy" offset is actually proportional to the pressure? |
|
I meant the offset from the perfect prediction line,not the absolute energy |
|
Oh, sorry @gabor1 . Probably they are proportional, these are from different training that I've done removing the not 0 pressure configurations.
|
|
Are you saying you get energy offsets even for all zero pressure data? The other thing that can cause it is incorrectly set E0 values when you have variable composition. Can you identify what characterises the configurations with each particular energy offset? |
|
Sure, but it will take me some time, I'll let you know. For all zero pressure data I get the good fitting. The other pictures are with the "good" dataset plus the same starting configurations and VASP inputs but just changing the PSTRESS value. One last thing, what do you mean with variable composition?? The dataset is made from different SiO2 structures. |
|
Ok ok, so you always have the same composition, that's good to know. So plot the prediction energy error against the external pressure you are applying, again separate colour for train and test |
|
Sorry what do you mean by "good fitting model"? Do you mean the one that was fitted on only zero pressure data? |
|
I'm interested to diagnose try he problems of your model that doesn't fit well. So please put the model trained on just zero pressure data aside, and plot the prediction error of the model which has the large energy errors, as a function of external pressure applied. |
|
Thanks. For each pressure, there are two groups of points: one group has small error, less than 10meV the other group much larger error, which grows with pressure. Can you look at the structures in each group and try to identify something that is different about them? Something that is unique to the groups with small error for example? |
|
Indeed! For both 1.0GPa and 1.5GPa groups, the points with small error are Single Point calculations. The points with bigger errors are from AIMD. I don't know if this is helpful or to be expected. |
|
This is almost conclusive evidence that the problem is your data. VASP must be giving data slightly different when it is doing AIMD and when doing single point. My bet is still that the +PV term is included in the AIMD data! So one easy solution (at a little bit of cost) is that since the AIMD trajectory is massively correlated anyway, you take some configs from it, and you recompute those with single point, exactly the settings you used for the points with small error. train only using single point data. |








Uh oh!
There was an error while loading. Please reload this page.
Hello! First of all I want to thank MACE team for all the work.
I have a question related with how finetuning a model deals with DFT calculations (made with VASP) with pressure. I've been obtaining a strange behaviour when I add configurations with PSTRESS=!0.0. In the second picture the configurations are the same than the first one (which are configurations with PSTRESS=0.0) plus with values of PSTRESS in [-0.4,-0.2,1.0,1.5]. I've finetuned MACE-matpes-r2scan model with r2scan+rvv10 data. This strange behaviour affects only the energy fitting.
Thank you in advance.
All reactions