ML-DFT: Machine learning for density functional approximations This repository contains the implementation for the kernel ridge regression based density functional approximation method described in the paper "Quantum chemical accuracy from density functional approximations via machine learning".

• python==3.7
• numpy>=1.19
• scikit-learn>=0.23
• scipy>=1.5
• sacred==0.8
• ase>=3.20

Clone the repository by running:

git clone https://github.com/MihailBogojeski/ml-dft.git

To install package along with all requirements, simply run:

cd ml-dft
pip install -r requirements.txt
pip install .

Install time should range from 3-10 minutes, depending on internet speed.

To train and test an example model based on a small dataset consisting of 102 water geometries, use the following command:

python run_ml_dft_experiment.py water_102 water_102 50 50

The training and evalutaion of this example model should be done in under a minute. The final lines of the output should look like:

Energies:
Correlation:  0.9999821536536118
RMSE:  0.15695036938603066
MAE:  0.08933831997099333
Max MAE:  0.6630057379879872

Since the crossvalidation process is random the values can be different, however if the model is trained successfully the MAE should be in the range from 0.2 to 0.06.