Improving the reliability of, and confidence in, DFT functional benchmarking through active learning

This repository contains the code and auxiliary data associated with the "Improving the reliability of, and confidence in, DFT functional benchmarking through active learning" project. Code is provided "as-is". Minor edits may be required to tailor the scripts for different computational systems.

Conda environment

To set up the conda environment: conda env create --name <env-name> --file environment.yml

Requirements

In order to execute the autodE high-throughput reaction profile computation workflow, Gaussian09/Gaussian16 and xTB needs to be accessible.

Curating the dataset

The scripts used for transforming the xyz-coordinates in reaction SMILES is scripts/analyze_data.py. Execution as follows:

python scripts/analyze_data.py --raw_data data/raw_data --iter initial --generate_initial_data

The xyz files should be in the directory data/XYZ_files. A data_smiles.csv file will be generated in the data directory. The script only works with neutral molecules, for the case of charged molecules, an error will be displayed, and you should add manually the reaction SMILES. The final version of the initial training set can be found in data/data_smiles_curated.csv

Generating the chemical space

The script used for generating the chemical space is script/generate_space.py. Execution as follows:

python scripts/generate_space.py --template_cores data/hypothetical_space_core.csv

A hypothetical_chemical_space.csv file will be generated in the data directory.

Baseline ML models

The script used for running the baseline models is script/baseline_models.py. The baseline_model.py script, which runs each of the baseline models sequentially, can be executed as follows:

python baseline_models.py --csv-file data/data_smiles_curated.csv

The fingerprints are generated during runtime. The DRFP and Morgan fingerprint is used. A nested cross-validation is implemented. For a final evaluation of the model you should use the option --final_cv

An example of the input file is included in the data directory: data_smiles_curated.csv.

Bayesian optimization campaign

Each iteration of the bayesian optimization campaign is launched with the help of the final_model.py script. Execution as follows:

final_model.py [-h] [--train_file TRAIN_FILE] [--pool_file POOL_FILE] [--new_data_file NEW_DATA_FILE] [--iteration ITERATION] [--seed SEED] [--beta BETA] [--final_dir FINAL_DIR] [--cutoff CUTOFF] [--conda_env CONDA_ENV] [--new_data NEW_DATA] [--selective_sampling SELECTIVE_SAMPLING] [--selective_sampling_data SELECTIVE_SAMPLING_DATA]

Reproducibility

To reproduce the acquired reaction each round as well the mean absolute error of each set and the plots, a bash script is provided. Execution as follows:

bash reproducibility.sh

An example of the autodE input can be found in data/autode_input_8.

References

If (parts of) this workflow are used as part of a publication please cite the associated paper: @article{ml_functional, author = {Alfonso-Ramos, Javier E. and Adamo, Carlo and Br{\'e}mond, {\'E}ric and Stuyver, Thijs}, title = {Improving the Reliability of, and Confidence in, {DFT} Functional Benchmarking through Active Learning}, journal = {J. Chem. Theory Comput.}, volume = {21}, number = {4}, pages = {1752-1761}, year = {2025}, doi = {10.1021/acs.jctc.4c01729}, }