Polymer Chemistry Informed Neural Networks (PCINNs) for Data-Driven Polymerization Modeling

This repository implements Polymer Chemistry Informed Neural Networks (PCINNs) for methyl methacrylate (MMA) free-radical polymerization prediction. The method combines data-driven training with kinetic-model guidance via Jacobian matching.

Paper: Ballard, N. Polymer Chemistry Informed Neural Networks. Polym. Chem., 2024. DOI: 10.1039/D4PY00995A
Upstream repository: https://github.com/PolymatGIQ/PCINN
Primary implementation: MMA_PCINN.ipynb

The original script MMA_PCINN.py is kept for reference only.

Quick Start

python -m venv .venv
source .venv/Scripts/activate   # Windows Git Bash
pip install torch==2.10.0 numpy==2.4.2 pandas==2.2.3 matplotlib==3.10.8 openpyxl==3.1.5 "ipykernel>=6.29,<7"

Open MMA_PCINN.ipynb in Jupyter/VS Code and run cells.

Notebook Workflows

The notebook contains two workflows:

1. Single-fold training/export workflow (quick path)

• Uses TestReaction to hold out one reaction (default: TestReaction = 8).
• Trains three models: Baseline NN, PCINN, Soft-Anchor PCINN.
• Plots loss curves and exports model artifacts.

2. Full leave-one-experiment-out multi-seed CV (long-running, optional)

• Runs all 8 folds and multiple seeds (N_SEEDS = 5 by default).
• Default run count: 3 models × 8 folds × 5 seeds = 120 training runs.
• Produces table-style aggregate metrics and fold-level PCINN vs SA-PCINN comparisons.

Notebook Structure

MMA_PCINN.ipynb currently has 30 cells (13 markdown, 17 code):

1. Dependencies and pinned install cell.
2. Imports/environment setup (including optional unsafe KMP gate, off by default).
3. Reproducibility and device setup (SEED=42, DEVICE auto-select).
4. Model definitions: NNmodel and DomainModel.
5. Data preprocessing and scaler/domain model loading.
6. Jacobian sampling range configuration.
7. Single-fold split configuration via TestReaction.
8. Baseline NN training (paper-original cell preserved as commented reference + improved executable cell).
9. PCINN training (paper-original cell preserved as commented reference + improved executable cell).
10. Soft-Anchor PCINN training.
11. Single-fold loss/curve comparison summary.
12. Full 8-fold multi-seed cross-validation section.
13. Model export with smoke-test reload verification.

Model Objectives

• Baseline NN: data-only MSE.
• PCINN: L_data + L_jac (Jacobian matching to pretrained theory model).
• Soft-Anchor PCINN: L_data + L_jac + λ_anchor * L_anchor, where:

$$L_{\text{anchor}} = \text{MSE}(y_{\text{net}}(x_j), y_{\text{theory}}(x_j))$$

using the same sampled points used for Jacobian matching.

Validation Checklist

Quick path:

• Restart kernel and run through the single-fold sections.
• Confirm all three model sections execute and render loss curves.
• Confirm NNpred, EBNNpred, and SApred each have shape (N, 6).
• Confirm export files are written to exports/ and smoke-test reload passes (max_diff < 1e-6).

Optional full CV:

• Run the multi-seed CV cell and confirm aggregate tables/plots are produced.
• Expect substantially longer runtime than the single-fold workflow.

License

This repository is licensed under CC BY-NC 4.0. See License.md.