Python code to reproduce all plots in:

❝Bayesian optimization of nanoporous materials❞ A. Deshwal, C. Simon, J. R. Doppa. Molecular Systems Design & Engineering. (2021) link preprint

requirements

the Python 3 libraries required for the project are in requirements.txt. use Jupyter Notebook or Jupyter Lab to run Python 3 in the *.ipynb.

search methods

step 1: prepare the data

our paper relies on data from Mercado et al. here. we visited Materials Cloud to download and untar properties.tgz giving properties.csv in new/. this is the data we use.

run the code in the Jupyter Notebook prepare_Xy.ipynb to prepare the data and write inputs_and_outputs.pkl to be read in by other Notebooks. in here, you can set the number of runs nb_runs, number of iterations for each run nb_iterations, and, if you wish, a flag downsample_data for testing.

step 2: run the searches

run the following Jupyter Notebooks, which will write search results to .pkl files.

• random_search.ipynb for random search
• evol_search.ipynb for evolutionary search (CMA-ES)
• random_forest_run.ipynb for one-shot supervised machine learning (via random forests). run twice, one with the flag diversify_training = True, the other with diversify_training = False.
• BO_run.ipynb for Bayesian optimization. run three times, with which_acquisition set to "EI", "max y_hat", and max sigma.

each .ipynb can be run on a desktop computer. the BO code takes the longest, at ~10 min per run.

step 3: visualize the results

finally, run viz.ipynb to read in the *.pkl files output from the search runs and visualize the results.

toy GP illustrations

see synthetic_example.ipynb for the toy GP plots in the paper.