GNN-PLUMED-Nucleation

Inputs for well-tempered metadynamics simulations biasing along GNN-learned reaction coordinates. Details can be found in this manuscript. Here we present a simplified procedure for constructing GNN RCs.

• Train and save the GNN model with Pytorch. The GNN models used in this work can be found in this github. For the two systems, we provided one ready-to-run model as model.pt for reproducing the molecular dynamics data.

• Convert the saved Pytorch model into libtorch version so PLUMED can take it as inputs. This is achieved via jit.trace function.

• Perform well-tempered metadynamics with MD engines patched with PLUMED. PLUMED needs to be compiled with PYTORCH_MODEL module. Instructions are provided in this colab tutorial and in PLUMED page.