cMBDF

convolutional MBDF via fast fourier transforms

cMBDF is also implemented in the qml2 library : https://github.com/qml2code/qml2

Usage similar to original MBDF (joblib) for entire dataset:

from cMBDF import generate_mbdf
reps = generate_mbdf(mols_charges, mols_coords, progress_bar=True) #to get a tqdm progress bar
reps_global = generate_mbdf(mols_charges, mols_coords, local=False) #to get flattened feature vectors

mols_charges and mols_coords should contain charges and coordinates arrays for multiple molecules to make this efficient.

Usage per molecule :

from cMBDF import get_convolutions, get_cmbdf

convs = get_convolutions() #only needs to be done once since the convolutions are unique, can also store it as .npy 

pad_size = max([len(q) for q in mols_charges]) #if not provided, defaults to molecule size with no padding
rep_list = [get_cmbdf(q,r,convs,pad_size) for q,r in zip(mols_charges, mols_coords)]

Note that the get_cmbdf function does not use parallelization. Also you can pass an arbitrarily large integer to pad_size if you don't know the largest molecule in your dataset. The results and timings are un-affected (except memory requirements due to redundant zeros).

Kernels and wrapper

• Wrappers for training KRR models using qmlcode and examples for using these scripts are available at : https://github.com/dkhan42/QMLwrap

References

Please cite following work :

1. Danish Khan, O. Anatole von Lilienfeld; Generalized convolutional many-body distribution functional representations. Proceedings of the National Academy of Sciences, 122 (41), e2415662122 (2025). https://doi.org/10.1073/pnas.2415662122

2. Danish Khan, Stefan Heinen, O. Anatole von Lilienfeld; Kernel based quantum machine learning at record rate: Many-body distribution functionals as compact representations. J. Chem. Phys. 21 July 2023; 159 (3): 034106. https://doi.org/10.1063/5.0152215