Physics-Guided Neural Networks for Transferable Prediction of Polymer Properties

Install Instructions

Please install using the setup.py.

$ git clone https://github.com/webbtheosim/gcgnn.git
$ cd gcgnn
$ conda create --name gcgnn python=3.8.16
$ conda activate gcgnn
$ # or source activate gcgnn
$ pip install -e .

Download Data and Results

Select a disk location for data storage and update the directory paths before running the program. Download the required data from Zenodo here.

• DATA_DIR: Stores data pickle files (approx. 210 MB).

Additional training weights and results can be downloaded here.

• TRAIN_RESULT_DIR or HIST_DIR (optional): Stores training results in pickle format (approx. 2.5 GB).
• MODEL_DIR (optional): Stores training weights in .h5 format (approx. 1.3 GB).

To train from scratch, only the DATA_DIR is required.

# LOAD SIMULATION DATA
DATA_DIR = "your/custom/dir/"
mw = 40 # or 90, 190 MWs

filename = os.path.join(DATA_DIR, f"pattern_graph_data_{mw}_{mw+20}_rg_new.pickle")
with open(filename, "rb") as handle:
    graph = pickle.load(handle)
    label = pickle.load(handle)
    desc  = pickle.load(handle)
    meta  = pickle.load(handle)
    mode  = pickle.load(handle)
    rg2_mean   = pickle.load(handle)
    rg2_std    = pickle.load(handle) ** 0.5 # var

# combine asymmetric and symmetric star polymers
label[label == 'stara'] = 'star'
# combine bottlebrush and other comb polymers
label[label == 'bottlebrush'] = 'comb' 

# LOAD GAUSSIAN CHAIN THEORETICAL DATA
with open(os.path.join(DATA_DIR, f"rg2_baseline_{mw}_new.pickle"), "rb") as handle:
    rg2_mean_theo = pickle.load(handle)[:, 0]
    rg2_std_theo = pickle.load(handle)[:, 0]

• graph: NetworkX graph representations of polymers. The node attribute "type" denotes the monomer type.
• label: Architectural classes of polymers (e.g., linear, cyclic, star, branch, comb, dendrimer).
• desc: Topological descriptors (optional).
• meta: Identifiers for unique architectures (optional).
• mode: Identifiers for unique chemical patterns (optional).
• rg2_mean: Mean squared radii of gyration from simulations.
• rg2_std: Corresponding standard deviation from simulations.
• rg2_mean_theo: Mean squared radii of gyration from theoretical models.
• rg2_std_theo: Corresponding standard deviation from theoretical models.

File Structures

To reproduce the figures and tables, refer to the notebook folder:

• figure_data.ipynb: Visualizes polymer architecture and chemical patterns used in the study.
• figure_result.ipynb: Visualizes all result figures from the paper.

The core codes are located in the gcgnn folder:

• train.py: Handles the training of GC, GNN, and GC-GNN models, including setup and execution.
• model_utils.py: Provides utility functions to support model training processes.
• gen_base_rg2.py: Generates theoretical mean and standard deviation values for Gaussian Chain squared radius of gyration squared.
• analysis: Contains scripts to calculate and evaluate accuracy metrics for model predictions.
• gen_data: Scripts for generating unique polymer architectures, represented as graph structures.
• gen_pattern: Assigns specific chemical patterns to the generated polymer architectures.
• model_data: Preprocesses datasets, including feature extraction and formatting, for model training.
• models: Defines the architecture and implementation of various machine learning models used in the study.
• plot: Scripts to create visual representations of data, including figures for analysis and publication.

The result_temp and csv_result folders contain temporary files used for figure preparation and table creation. To generate your own results, set rerun=True in notebooks.

The submit folder contains job submission scripts tailored for High-Performance Computing environments.

• submit_train.submit: Train neural networks.
• gen_base_rg2.submit: Generate all theoretical Rg predictions.