LFaB-for-QC

Code repo for study involving use of low fidelity as bias in active learning sampling strategy for training data sampling in chemical configuration space. The scripts in this repository can be used on the datasets used in the work. The paper can be accessed at https://arxiv.org/abs/2508.15577 as a preprint. If the interest is to only reproduce the plots, use the jupyter notebook.

Software Setup

In order to execute the experiments performed in this work, it is adviced that you set up a fresh environment with certain pythonic libraries installed. The following steps will be of relevance.

$ conda create --name LFaB_env python=3.9.18

Activate the environment with

$ conda activate LFaB_env

NOTE: If installing the qml package, follow steps from https://github.com/vivinvinod/QeMFi/tree/main.

Install the required packages within this environment using:

(LFaB_env)$ pip install -r requirements.txt

This environment can now be used to run the scripts from this code repository.

Datasets Used

1. QM7b - Montavon, G., et al. https://doi.org/10.1088/1367-2630/15/9/095003
2. VIB5 - Zhang, L., et al. https://doi.org/10.1038/s41597-022-01185-w
3. QeMFi - Vinod, V., et al. https://doi.org/10.1038/s41597-024-04247-3

It is assumed that the data is downloaded into directories named QM7b/, VIB5/, and QeMFi/ respectively.

Once downloaded, the next step is to generate molecular descriptors. In this work, two molecular descriptors are used:

Representations Used

1. Unsorted Coulomb Matrices - Rupp, M., et al. https://doi.org/10.1103/PhysRevLett.108.058301
2. SLATM - Huang, B., et al. https://doi.org/10.1038/s41557-020-0527-z

Scripts to generate these descriptors for a given dataset are of the general form provided in GenerateCM.py and GenerateSLATM.py respectively. The correct path variable and number of samples can be changed as needed. For QeMFi, scripts from https://github.com/vivinvinod/QeMFi/tree/main can be used for representation generation.

Reproducing the Results

All plotting functions to produce the figures from the manuscript are found in plottingroutines.ipynb. These load pre-calculated data stored in ModelData/ and PlotData/. At the same time, users can run their own computations using the scripts provided. Scripts labelled GPR_AL_<dataset>.py correspond to each dataset and can be run with modification to the path of the generated molecular descriptors. The file GPR_model.py is a simply GPR implementation using GpyTorch. For the QeMFi dataset, the script can be run specifying the molecule name directly from the terminal as

(LFaB_env)$ python GPR_AL_QeMFi.py sma

which will produce the results for the SMA molecule from the dataset.