-
|
I noticed that the MACE-OFF23 was originally trained on a subset of SPICE that only contained neutral molecules. Was there a reason for the exclusion of charged molecules and are there any modifications that are required to support training on both neutral and charged molecules? |
Beta Was this translation helpful? Give feedback.
Replies: 1 comment 4 replies
-
|
yes, the currently published MACE model has no explicit analytic coulomb interactions. You can of course train it on the entire spice dataset including charged systems, but it will be much less accurate, because the different energy of charged and uncharged systems is poorly reflected in just the geometry only. We are working on adding explicit charges to MACE, at which point the entire database will be refitted. Not far away now (it works for us in-house already, and is undergoing testing before release) |
Beta Was this translation helpful? Give feedback.
-
|
Thank you for the prompt reply. |
Beta Was this translation helpful? Give feedback.
-
|
of course there are many ways in one can go about this. We are using a charge-self-consistent approach, in which the atom energy is a function of local geometry and field, and this is iterated to self-consistency. |
Beta Was this translation helpful? Give feedback.
-
|
Hi, will this function be published any soon? |
Beta Was this translation helpful? Give feedback.
-
|
months I would say. |
Beta Was this translation helpful? Give feedback.
yes, the currently published MACE model has no explicit analytic coulomb interactions. You can of course train it on the entire spice dataset including charged systems, but it will be much less accurate, because the different energy of charged and uncharged systems is poorly reflected in just the geometry only. We are working on adding explicit charges to MACE, at which point the entire database will be refitted. Not far away now (it works for us in-house already, and is undergoing testing before release)