Skip to content

reactive_atom

Directory actions

More options

Directory actions

More options

Latest commit

 

History

History

reactive_atom

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
README.md
README.md
 
 
data.py
data.py
 
 
loss.py
loss.py
 
 
model.py
model.py
 
 
train.ipynb
train.ipynb
 
 

README.md

GNN for Selecting Reactive Atoms

Predict reactive atoms

Implemented

  • MPNN, GraphSAGE, GAT models
  • Energy head (atomwise decomposition)
  • Localization head (per-orbital logits --> softmax over atoms)
  • Loss: 
    Loss = Energy MSE + optional force loss + -beta * J_sur + orthogonality regularizer
  • Training loop with increase of localization weight (β)
  • Validation routine that calculates Pipek-Mezey (PM) via PySCF:
    • energy MAE
    • force MAE (if force provided)
    • surrogate J_sur v.s. true J_PM correlation
    • per-orbital per-atom AUC (ranking) using top-k ground-truth labeling
  • Can save/load checkpoints