A Model Ensemble Approach Enables Data-Driven Property Prediction for Chemically Deconstructable Thermosets in the Low-Data Regime

Yasmeen AlFaraj, Somesh Mohapatra, Peyton Shieh, Keith E. L. Husted, Douglass G. Ivanoff, Evan M. Lloyd, Julian C. Cooper, Yutong Dai, Avni Singhal, Jeffrey S. Moore, Nancy R. Sottos, Rafael Gómez‐Bombarelli, Jeremiah A. JohnsonACS 2023 · DOI

TL;DR

This paper introduces a data-driven approach using a machine learning model ensemble to predict the glass transition temperature (Tg) of deconstructable thermosets, even with limited experimental data. This method accelerates the discovery of new materials with tunable properties and improved sustainability by enabling accurate predictions of how molecular building blocks influence material performance.

What Problem Does This Paper Address?

Thermosets are widely used for their durability and resistance to extreme conditions, but their cross-linked nature makes them difficult to recycle or deconstruct, posing a significant sustainability challenge. While deconstructable thermosets offer a promising solution, the sheer number of possible molecular components (monomers, cross-linkers, additives) and manufacturing conditions creates a vast combinatorial space. Predicting how these components translate into bulk material properties, like the glass transition temperature (Tg), is crucial for designing materials with desired performance and end-of-life options. However, traditional computational methods struggle with the complex, amorphous, and often low-data nature of these multi-component polymer networks.

Methods

The researchers developed a "closed-loop experimental, machine learning (ML), and virtual screening strategy." Key to their approach is the use of a model ensemble. Instead of relying on a single ML model, they combined predictions from multiple models with varying architectures and hyperparameters. This ensemble approach, coupled with heavy regularization, helps to mitigate overfitting, especially in low-data regimes, and provides a measure of prediction uncertainty.

For input, they represented the chemical precursors (cleavable comonomers and cross-linkers) as small molecules using low-dimensional physicochemical features and extended connectivity fingerprints. This avoids the need to explicitly represent the complex 3D network topology of the thermoset. The model was trained to predict the glass transition temperature (Tg) based on these molecular features and formulation variables (composition and loading of additives).

Datasets

The study utilized a curated dataset of 101 experimental data points for training and evaluating their ML model. This dataset was compiled from existing literature and in-house experiments, focusing on polydicyclopentadiene (pDCPD) thermosets incorporating cleavable bifunctional silyl ether (BSE) comonomers and/or cross-linkers. The paper also includes a retrospective analysis using a different class of previously reported cleavable additives (strand-cleaving cross-linkers, SCCs) that were not part of the original training data, serving as an additional evaluation of the model's generalization capabilities.

Key Insights and Ideas

  • Ensembling for Low-Data Regimes: The core innovation is the successful application of model ensembling to achieve robust predictions for complex materials with limited data. This approach effectively balances bias-variance trade-offs, leading to higher accuracy and more reliable uncertainty quantification than single models.
  • Feature Representation for Amorphous Polymers: Representing individual molecular precursors with physicochemical features and fingerprints, rather than attempting to model the entire network structure, proved effective for capturing relevant chemical information without requiring explicit topological data.
  • Predicting Deconstructability and Performance: The model not only predicts a critical performance metric (Tg) but also enables the design of thermosets that are both deconstructable and possess tunable thermal properties, offering a pathway to more sustainable materials.
  • Identifying Key Structure-Property Relationships: The analysis of feature importance revealed that cleavable comonomer/cross-linker concentration, initiator type, and specific molecular substructures (like aryl substituents and norbornene-based cross-linkers) significantly influence Tg.

Key Results

The model ensemble achieved a prediction accuracy of within ±15 °C for the glass transition temperature (Tg) of deconstructable pDCPD thermosets. This was demonstrated through:

  • High-fidelity prediction: A parity plot showed good agreement between predicted and experimental Tg values for the training data, with an RMSE of 14.88 °C.
  • Successful validation of new materials: The model accurately predicted the Tg of novel BSE-based cleavable comonomers (LinF7 and PhSi7) that were not included in the training set, with experimental values falling within the predicted uncertainty.
  • Retrospective prediction success: The model also showed reasonable agreement in predicting Tg for a different class of cleavable additives (SCCs) that were entirely absent from the training data, highlighting its generalization potential.
  • Deconstructability confirmed: Synthesized materials based on model predictions were experimentally confirmed to be chemically deconstructable, dissolving into soluble fragments.

Limitations and Future Work

  • Focus on Tg: The current model is specifically designed to predict the glass transition temperature (Tg). The authors suggest that extending this framework to predict other thermomechanical properties (e.g., elastic modulus, yield stress, cross-link density) would be a valuable future direction.
  • Environmental Considerations: While the model predicts desirable material properties, it does not inherently consider environmental impacts, such as the potential release of perfluoroalkylated substances (PFAS) from certain cleavable comonomers. Future models could incorporate such considerations, along with cost, into the prediction process.
  • Data Scarcity: Although the ensemble approach is effective for low-data regimes, the authors acknowledge that more data would further improve accuracy and enable the incorporation of more complex features and properties.
  • Retrospective Analysis Nuances: For one specific composition in the retrospective analysis (100% v/v % NbMeSi thermoset lacking DCPD), the model's prediction was inaccurate, and uncertainty was high. This highlights that predictions are most reliable when the material composition is within the scope of the training data (i.e., pDCPD-based).

Summary updated

Abstract

Thermosets present sustainability challenges that could potentially be addressed through the design of deconstructable variants with tunable properties; however, the combinatorial space of possible thermoset molecular building blocks (e.g., monomers, cross-linkers, and additives) and manufacturing conditions is vast, and predictive knowledge for how combinations of these molecular components translate to bulk thermoset properties is lacking. Data science could overcome these problems, but computational methods are difficult to apply to multicomponent, amorphous, statistical copolymer materials for which little data exist. Here, leveraging a data set with 101 examples, we introduce a closed-loop experimental, machine learning (ML), and virtual screening strategy to enable predictions of the glass transition temperature (<i>T</i><sub>g</sub>) of polydicyclopentadiene (pDCPD) thermosets containing cleavable bifunctional silyl ether (BSE) comonomers and/or cross-linkers with varied compositions and loadings. Molecular features and formulation variables are used as model inputs, and uncertainty is quantified through model ensembling, which together with heavy regularization helps to avoid overfitting and ultimately achieves predictions within <15 °C for thermosets with compositionally diverse BSEs. This work offers a path to predicting the properties of thermosets based on their molecular building blocks, which may accelerate the discovery of promising plastics, rubbers, and composites with improved functionality and controlled deconstructability.
BibTeX
@inproceedings{alfaraj2023model,
  author = {Yasmeen AlFaraj and Somesh Mohapatra and Peyton Shieh and Keith E. L. Husted and Douglass G. Ivanoff and Evan M. Lloyd and Julian C. Cooper and Yutong Dai and Avni Singhal and Jeffrey S. Moore and Nancy R. Sottos and Rafael Gómez‐Bombarelli and Jeremiah A. Johnson},
  title = {{A Model Ensemble Approach Enables Data-Driven Property Prediction for Chemically Deconstructable Thermosets in the Low-Data Regime}},
  booktitle = {ACS},
  year = {2023},
  doi = {10.1021/acscentsci.3c00502},
}
Figure