search

isayevlab / AIMNet2

MIT License
100 stars 25 forks source link

Inconsistency between old model and new model (specifically for charges) #43

Closed peterTHC closed 1 month ago

peterTHC commented 1 month ago

Hi,

I was trying to use AIMNET2 to calculate energies and obtain reaction profile. I have used the old model and aimnet2ase.py to calculate energies of a substitution reaction. Recently, I changed to use the new model. I used exactly the same method but got different results. The system has a overall charge of -1. For the old model, I was able to obtain the transition state using set_charge(-1). However, when I tried the new model, I got similar results using set_charge(0) and set_charge(-1) and the transition state is a lot harder to converge. I checked my optimized reactant finding that both charge 0 and charge -1 give me similar geometry while I would expect -1 will give me a more reasonable geometry. I am wondering if there could be something changed for the charge information that should be passed to the neural network.

The substitution I am trying is FCCN+=O.[Cl-]>>ClCCN+=O.[F-]

Thank you!

peterTHC commented 1 month ago

I've figured out another way to finalize my energy profile. But I still found the optimization of FCC[N+]=O.[Cl-] is a little weird. The chloride is too close to the hydrogens on the carbon. Not sure if the two-fragment case is harder.