Repository navigation
Replies: 4 comments 8 replies
|
what are the two groups of points on the energy parity plot? |
|
are these structures all relaxed? because the force errors are still very small (even though I obviously see the problem with the lack of correlation). generally we like to train models with rattled configurations, so that you get a good range of forces. all your forces are quite small. |
|
how did you select the 0.25 A^-1 ? do you know that it's actually converged in terms of energies and forces to better than 1meV for the energy and 0.01 eV/A for forces? |
|
Thanks for the previous feedback and I have recomputed DFT with kpoint density 0.15 A^-1 (previously 0.25 A^-1) and now the error has been reduced compared to before. However, still I have a bit of concern on the overall accuracy when I have to deal with large system. The overall RMSE is quite low like RMSE 0.5 meV/atom and 6.5 meV/A but since I'm dealing with large system with number of atoms from 500 - 1500, when it's multiplied with the number of atom, the error can be as large as 1-2 eV. Also when I look at parity plot of slab systems, (without adsorbates) it's more or less flat which is concerning. The energy difference is very small so still RMSE is 0.6 meV/atom but scatters are not aligned to parity line. I wonder if this is still DFT convergence problem or is it sort of an accuracy that is hard to achieve? |



Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
Hello, I'm trying to naively finetune MACE from mace-mh1 model for various Au111 slabs.
My training set of Au111 have various supecell sizes 4 layers with number of atoms ranging 560 - 1200 atoms.
Because slab structures in training siet have different supercell, I used kspacing desnity of 0.25 A-1 for DFT calculation in FHI-aims.
Following is parity plot for energy/atom and forces. I intentionally evaluated this parity plot for the training set just to see whether training is properly done. However, even for prediction of its own training set, it seems to be failing. (Bottom plot is zoom in for red area in the upper left plot)

I'm using naive fine-tuning instead of multihued fine-tuning because I wanted to to check whether training works for this system.
Following is my naive fine-tuning keywords:
I have attached five example structures of training sets just to give you idea of how does my training set look like.
select.xyz.zip
I'm a bit puzzled why this training on metal slab models doesn't work.
Is it related to inconsistent DFT calculation that is beyond kspacing issue?
Or is it too challenging to train a MACE for this various supercell sizes?
I'll be appreciated for any comments and if you need further information to decipher this, please let me know!
Many thanks in advance.
All reactions