By how much can closed-loop frameworks accelerate computational materials discovery?

Data and scripts in support of the publication "By how much can closed-loop frameworks accelerate computational materials discovery?", Kavalsky et al., arXiv:2211.10533 [cond-mat.mtrl-sci] (2022). DOI: 10.48550/arXiv.2211.10533.

The repository is organized as follows:

1. data/

   • benchmark_calculations_record.xlsx: Excel spreadsheet containing a record of DFT calculations, associated raw timestamps, and a tabulation of the acceleration estimates.

   • bimetallic_catalysts_dataset/

     • ma_2015_bimetallics_raw.json.gz: Dataset of bimetallic alloys for CO2 reduction, in the Physical Information File (PIF) format, obtained from Dataset 153450 on Citrination.

       Original data source: "Machine-Learning-Augmented Chemisorption Model for CO2 Electroreduction Catalyst Screening", Ma et al., J. Phys. Chem. Lett. 6 3528-3533 (2015). DOI: 10.1021/acs.jpclett.5b01660

     • transform.py: Python script for converting from the PIF format into tabular data.

     • bimetallics_data.csv: Bimetallics catalysts dataset mentioned above in a tabular format.

   • runtime_geometries/

     "Chemically-informed" and naive structures and settings in the form of ase.traj files, corresponding to the discussion surrounding FIG. 3 in the main text. The files can be read using ASE package (using ase.io.read).

2. scripts/

   • human_lagtime.py: Script for estimating human lagtime in job management, calculated using a Monte Carlo sampling method.

   • sequential_learning.py: Script for running multiple independent trials of sequential learning (SL) and recording a history of training examples, model predictions and prediction uncertainties.

     If run as-is, the script performs 20 independent trials of 100 SL iterations to optimize the binding_energy_of_adsorbed property in the bimetallic catalysts dataset mentioned above, using four acquisition functions (results from each recorded separately): random, maximum likelihood of improvement (MLI), maximum uncertainty (MU), and space-filling.

   • plot_acceleration_from_sequential_learning.py: Script to aggregate results from the sequential_learning.py script, calculate and plot statistics related to acceleration from SL over a baseline.

     If run as-is, the script reproduces the 3-paneled FIG. 5 in the main text.

   • plot_acceleration_from_sequential_learning__ALL_ACQ.py: Similar to the previous script; plots and compares statistics from all acquisition functions considered (MLI, MU, random, space-filling) for all SL tasks.

     If run as-is, the script reproduces the 3-paneled FIG. S1 in the Supplementary Information.

   • plot_levels_of_automation.py: Script for plotting the cumulative time of executing the DFT pipeline at varying levels of automation.

     If run as-is, the script reproduces the bottom panel from FIG. 2 in the main text.

Running the scripts

The required packages for executing the scripts are specified in requirements.txt, and can be installed in a new environment (e.g. using conda) as follows:

$ conda create -n accel_benchmarking python=3.10
$ conda activate accel_benchmarking
$ pip install -r requirements.txt

The scripts are all in python, and can be run from the command line. For example:

$ cd scripts
$ python sequential_learning.py