Functional monomer design for synthetically accessible polymers

This repository contains python scripts used to obtain comprehensive monomer-level properties for synthetically accessible Open Macromolecular Genome (OMG) polymers. as described in the paper. Trained ML models, conformer geometry, and ML-based monomer-level properties for 12M OMG polymers with prediction uncertainties are available at Zenodo.

Set up Python environment with Anaconda

conda env create -f environment.yml

Script components

To run a script, a file path in the script should be modified to be consistent with an attempted directory.

1. active_learning

This directory contains scripts to perform active learning campaign based on evidential learning. To use trained models, download pareto_greedy_check_point.tar.gz from Zenodo and extract pareto_greedy_check_point.tar.gz with the following command:

tar -xvzf pareto_greedy_check_point.tar.gz

The current file path in the scripts assume that the pareto_greedy_check_point directory is located at ./active_learning

2. data

This directory contains quantum chemistry calculation results during the active learning campaign. For example, ./active_learning/pareto_greedy/OMG_train_batch_0_chemprop_with_reaction_id.csv contains quantum chemistry calculation results from the initial train data. Similarly, ./active_learning/pareto_greedy/OMG_train_batch_1_chemprop_with_reaction_id.csv contains quantum chemistry calculation results from Round 1. This directory also contains 200 RDKit features (./rdkit_features) used in training ML models to overcome the local nature of message passing in graph neural networks.

The examples contains one example reaction of quantum chemistry calculations for different conformers. The full data can be downloaded at Zenodo.

3. experimental_chi

This directory contains scripts to estimate Flory-Huggins interaction parameters from COSMO-SAC calculations.

4. external_packages

This directory contains external packages including D-MPNN networks and a Pareto front search algorithm from GitHub.

5. figure_publication

This directory contains scripts to draw figures in the paper.

6. qm9_active_learning_add_100 / qm9_active_learning_add_1000

This directory contains example scripts to perform active learning on QM9 for the figures in Supporting Information.

7. run_calculation

This directory contains scripts to run quantum chemistry calculations.

• batch_conformer_FF.py (conformer search and geometry optimization with UFF)
• batch_conformer_xtb.py (geometry optimization with XTB2)
• batch_rdkit_dft_property_calculation.py (DFT and TD-DFT calculations)
• batch_save_dft_chi_results.py (save DFT results and calculate Flory-Huggins interaction parameters)
• batch_save_total_results.py (save results)

8. utils

This directory contains functions needed to run scripts.

Authors

Seonghwan Kim, Charles M. Schroeder, and Nicholas E. Jackson

Funding Acknowledgements

This work was supported by the IBM-Illinois Discovery Accelerator Institute. N.E.J. thanks the 3M Nontenured Faculty Award for support of this research.