Graph Convolutional Neural Networks for Polymers Property Prediction

Minggang Zeng, Jatin Kumar, Savitha Ramasamy, Vijay Chandrasekhar, Kedar HippalgaonkararXiv 2018 · DOI · arXiv

Abstract

A fast and accurate predictive tool for polymer properties is demanding and will pave the way to iterative inverse design. In this work, we apply graph convolutional neural networks (GCNN) to predict the dielectric constant and energy bandgap of polymers. Using density functional theory (DFT) calculated properties as the ground truth, GCNN can achieve remarkable agreement with DFT results. Moreover, we show that GCNN outperforms other machine learning algorithms. Our work proves that GCNN relies only on morphological data of polymers and removes the requirement for complicated hand-crafted descriptors, while still offering accuracy in fast predictions.

Introduction

Designing new polymers for electronics or energy storage usually requires knowing their physical properties in advance. Traditionally, scientists use Density Functional Theory (DFT)—a gold-standard simulation method—to calculate these properties. However, DFT is computationally expensive and slow, making it impractical for screening thousands of potential new materials.

Previous attempts to use machine learning (ML) to speed this up relied on "hand-crafted descriptors" (like SMILES strings or QSPR features). These are manual summaries of a molecule's structure that often fail to capture the complex, three-dimensional "morphological character" of a polymer. Furthermore, polymers are often amorphous, making them much harder to model than simple crystals.

Methodology

The authors propose using Graph Convolutional Neural Networks (GCNNs). Instead of a human telling the model which chemical features are important, the GCNN treats the polymer's repeat unit as a mathematical graph where atoms are nodes and chemical bonds are edges.

Schematic of the Graph CNN architecture showing the transition from 3D morphological data to property prediction.

The model uses "graph convolutions" to allow each atom to "talk" to its neighbors, automatically learning the local chemical environment. The architecture consists of:

  1. Input: Crystallographic Information Files (CIF) converted into 2D graphs.
  2. Convolutional Layers: Multiple layers that iteratively update atomic features based on surrounding bonds and atoms.
  3. Pooling: A normalized summation layer that collapses the graph into a single "fingerprint" vector for the whole polymer.
  4. Fully Connected Layers: A regression head that predicts the final property (dielectric constant or bandgap).

Data and Experiments

The study utilized the Polymergenome Project dataset, containing 1,073 polymers categorized into three subsets: Experimental (34 common polymers), COD Subset (253 organic/organometallic polymers), and a Computational Subset (786 polymers generated via building blocks). The dataset was split into a 6:2:2 ratio for training, validation, and testing.

The GCNN approach was compared against four traditional ML algorithms (Kernel Regression, Random Forest, Gradient Boosting, and standard Neural Networks) which were fed 158 manually extracted features using the Matminer library.

Results and Discussion

The GCNN achieved "remarkable agreement" with DFT results, consistently yielding the lowest Mean Absolute Error (MAE) across the board.

Performance of GCNN in predicting dielectric constant and bandgap compared to DFT ground truth, along with property distributions.
Property GCNN (Proposed) Best Traditional ML (Method)
Dielectric Constant ($\epsilon$) 0.24 0.355 (Random Forest)
Bandgap ($E_g$) 0.41 eV 0.446 eV (Gradient Boosting)

Key insights include:

  • Morphology is Critical: For ground-state properties like dielectric constants, the local arrangement of atoms (morphology) is the most critical factor, which GCNNs capture naturally.
  • Feature Learning vs. Engineering: The GCNN outperformed models using 158 hand-crafted features, suggesting that "latent" features learned by the network are more representative of the underlying physics than descriptors humans traditionally choose.
  • Data Scaling: Error systematically decreased as dataset size increased, indicating potential for further improvement with more data.

Conclusion

While the model shows high accuracy, it is currently trained on DFT-calculated values rather than experimental data, and focuses on bulk properties of pure polymers. Future work will aim to improve interpretability to understand why the GCNN makes certain predictions, potentially providing new chemical insights for inverse design.

Summary updated

Abstract

A fast and accurate predictive tool for polymer properties is demanding and will pave the way to iterative inverse design. In this work, we apply graph convolutional neural networks (GCNN) to predict the dielectric constant and energy bandgap of polymers. Using density functional theory (DFT) calculated properties as the ground truth, GCNN can achieve remarkable agreement with DFT results. Moreover, we show that GCNN outperforms other machine learning algorithms. Our work proves that GCNN relies only on morphological data of polymers and removes the requirement for complicated hand-crafted descriptors, while still offering accuracy in fast predictions.
BibTeX
@misc{zeng2018graph,
  author = {Minggang Zeng and Jatin Kumar and Zeng Zeng and Savitha Ramasamy and Vijay Chandrasekhar and Kedar Hippalgaonkar},
  title = {{Graph Convolutional Neural Networks for Polymers Property Prediction}},
  year = {2018},
  eprint = {1811.06231},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1811.06231},
}
Figure