Closed WardLT closed 8 months ago
We have previously tested this, the success rate on hMOF dataset is not very high.
Interesting! What do you use instead? Is it the code inside the "DiffLinker" module?
For DiffLinker, the inputs of the training data (ligands) are SMILES strings, so we borrowed MOFid from the same group published hMOF. The success rate is higher than the MOF fragmentor.
We also tried MOFSimplify from Dr. Heather Kulik's group, but there are other problems, like: ligands crossing the PBC are not recognized as one whole ligand, large chunk of the MOF structure with both metal and organic ligands was identified as a node, etc.
We have another generative model (GAN) which takes 3D coordinates of the ligands. MOFid has problems producing isolated ligand structures with 3D coordinates.
We also have developed our own code to deal with fragmentation of MOFs, which attempts to solve the known problems from other software tools.
I integrated 2 new modules into our repo: RASPA2 and PACMOF. RASPA2 is a Monte Carlo code using classical force field for CO2 adsorption simulation. PACMOF is a scikit-learn-based machine learning code to assign partial charges to MOF structures. It is faster than the charge equilibrium method in RASPA2, so probably worth trying.
RASPA2 (https://github.com/iRASPA/RASPA2) and edited the source code to include a new type of angle interaction (https://docs.lammps.org/angle_fourier.html). The forked the source code: https://github.com/williamyxl/RASPA2. Perparation code for RASPA: https://github.com/globus-labs/mof-generation-at-scale/blob/add_cif2lammps/mofa/simulation/raspa.py.
Also for PACMOF: https://github.com/arung-northwestern/pacmof, I integrated a modified version into our repo: https://github.com/globus-labs/mof-generation-at-scale/blob/add_cif2lammps/mofa/simulation/pacmof.py
We no longer need this reference
There are