datagen

Subpackage for Kohn-Sham calculations, density fitting and training data generation.

The subpackage provides all necessary functionalities to translate molecular geometry datasets into neural network training data.

Overview

The subpackage is divided into two parts: low-level modules and high-level modules. The low-level modules work on single molecules and do all the necessary calculations. The high-level modules scale these calculations to datasets and ensure correct saving of the data as well as parallelization.

Low-level Modules (Methods)

mldft.datagen.methods.ksdft_calculation wraps and patches pyscf to save every iteration of the Kohn-Sham computation.
mldft.datagen.methods.density_fitting fits coefficients of a new basis to the coefficients of the Kohn-Sham basis.
mldft.datagen.methods.label_generation calculates labels for energies and gradients.
mldft.datagen.methods.save_labels_in_zarr_file saves computed labels in a .zarr file.

High-Level Modules

mldft.datagen.kohn_sham_dataset handles Kohn-Sham calculations on a dataset.
mldft.datagen.generate_labels_dataset handles density fitting, label generation and saving of labels on a dataset.

Datasets

mldft.datagen.datasets.dataset defines the interface for a dataset.
mldft.datagen.datasets.qm9 provides the QM9 dataset.
mldft.datagen.datasets.qmugs provides the QMUGS dataset.
mldft.datagen.datasets.misc provides a general dataset for arbitrary xyz files.

Current Timings

The current timings on large machines using 64 cores, 128 threads in parallel.

QM9 Kohn-Sham: 8 seconds per iteration -> 300 CPU hours
QM9 Density Fitting: 10 seconds per iteration -> 360 CPU hours

Modules

`datasets`	Contains the data generation dataset classes.
`generate_labels_dataset`	Compute labels for molecules in the dataset and save them as zarr.zip files.
`kohn_sham_dataset`	Computing Kohn-Sham data for a dataset.
`methods`	Methods for Kohn-Sham calculations, density fitting and label saving.
`transform_dataset`	Applying a basis transformation to the dataset.