-
|
Hi all, I'd appreciate your expertise on the following: I am aware that MACE does not explicitly consider long-range charge effects but that short-range charge effects can be modeled quite well with MACE. I am also aware that retraining MACE is performed on absolute energies and forces, making it imperative that any new data be provided with the same zero of energy. I would like to retrain MACE on two dissociation curves: one of a cationic molecule dissociating from a neutral molecule, and one of a different molecule in anionic form dissociating from a neutral molecule. I think it should be okay to include the dissociation curve with charged species into the training data alongside the SPICE dataset, as we could have an equivalent system where the PFAS was protonated, and then the proton is removed to a distance of >10 Å (i.e. beyond the interaction length of MACE). As the energy of the isolated proton is zero within the Born-Oppenheimer approximation, I don't believe its exclusion will change the zero of energy that is important to preserve for a consistent training set. Is this assessment correct? Also worth noting that the training calculations will be done in gas phase, I suppose I could use CPCM if the charge screening is an issue. Thank you! |
Beta Was this translation helpful? Give feedback.
Replies: 1 comment 2 replies
-
|
The problem is about inconsistent data where the geometries are very close but the charge state (and hence the energy) is very different. So an isolated proton has very different energy from the isolated H, but geometrically the same. That's the problem. Electrostatic mace will solve it in a few months. In the meantime, you could try condensed phase training with some charged species, because there might be sufficient geometric information there to learn the energies (and implicitly the charge state). But I wouldn't mess with the isolated proton, that's really inconsistent. |
Beta Was this translation helpful? Give feedback.
-
|
If you really want to experiment, you can retrain (but not finetune!) from scratch taking samples from SPICE and your own data, but give the isolated atom energies from calculations where you ONLY include a proton (i.e. E=0) rather than H (E=-13.6). that's at least consistent. Just don't ever add H2 to your dataset :-) |
Beta Was this translation helpful? Give feedback.
-
|
I see, thank you for the detailed response. I'll try with explicit solvent and eagerly await electrostatic MACE! |
Beta Was this translation helpful? Give feedback.
The problem is about inconsistent data where the geometries are very close but the charge state (and hence the energy) is very different. So an isolated proton has very different energy from the isolated H, but geometrically the same. That's the problem. Electrostatic mace will solve it in a few months. In the meantime, you could try condensed phase training with some charged species, because there might be sufficient geometric information there to learn the energies (and implicitly the charge state). But I wouldn't mess with the isolated proton, that's really inconsistent.