Gaussian Integral (GI) Descriptors

Instruction

1. Run Gaussian Approximation

To perform Gaussian approximation of the original Lennard-Jones (LJ) potential curves, run:

python gauss_approx_lj.py

To perform Gaussian approximation of the soft LJ curves, run:

python gauss_approx_soft_lj.py

The optimized Gaussian parameters for each element will be saved in:

• gauss_params/lj for the origianl LJ potential
• gauss_params/soft_lj for the soft LJ potential

2. Generate GI Descriptor Vectors

To compute the GI descriptors, run the following command:

python gi.py --system --name OPAGIX --sigmas 0.1 0.2 0.3 --mcsh-orde 2 --mof-pool mean --mol-pool com

Required Arguments

• --system: specifies the system type
  • mof: for metal-organic frameworks
  • mol: for molecules
• --name: name of the compound
  • If --system is mof, it will read mof/{name}.cif
  • If --system is mol, it will read mol/{name}.xyz
• --sigmas: space-separated list of sigma values used to generate the GI descriptors
• --mcsh-order: Maximum MCSH order to include in the descriptor
• --mof-pool: pooling strategy for MOFs
  • min: use the minimum value across the structure
  • mean: use the mean value across the structure
• --mol-pool:
  • mean: use the mean value across all atoms
  • com: compute a single descriptor vector at the center-of-mass

The resulting GI descriptors will be saved in the descriptors/ directory with the filename {system}.csv.

Dependencies

• NumPy
• SciPy
• Atomic Simulation Environment (ASE)
• PyTorch
• Skorch
• AMPTorch

---

If you use the GI descriptor scheme in a scientific publication, please cite the following paper:

S. Choi, D. S. Sholl, and A. J. Medford, Gaussian Approximation of Dispersion Potentials for Efficient Featurization and Machine-Learning Predictions of Metal-Organic Frameworks, J. Chem. Phys. 2022, 156, 214108. DOI: https://doi.org/10.1063/5.0091405

---

Acknowledgements

• This works was supported by the Department of Energy, Office of Science, Basic Energy Sciences, under Award #DE-SC0020306.
• The codes in gi.py is adapted from AMPtorch/CEMT. For installation and instruction of the AMPtorch package, please refer to its official GitHub repo.