TL;DR
This paper introduces a machine learning (ML) model that can rapidly predict the ring-opening polymerization enthalpy (ΔH<sup>ROP</sup>) for polymers. By fusing experimental data with computationally generated data, the model achieves an accuracy of ~8 kJ/mol, significantly accelerating the design of new, recyclable polymers.
What Problem Does This Paper Address?
The growing problem of plastic waste highlights the urgent need for more sustainable materials. Recyclable polymers, particularly those formed via ring-opening polymerization (ROP), offer a promising solution. A key thermodynamic property governing the feasibility of ROP and its reverse (depolymerization) is the ring-opening enthalpy (ΔH<sup>ROP</sup>). While experimental measurements are the gold standard, they are time-consuming and limited in scope. First-principles computations can generate more data, but they are computationally expensive, hindering rapid exploration of new polymer chemistries. This paper addresses the challenge of efficiently predicting ΔH<sup>ROP</sup> to guide the design of depolymerizable polymers by developing a fast and accurate ML model.
Methods
The authors employed a multitask machine learning approach to leverage both experimental and computational data. This strategy exploits correlations between different data sources to improve model robustness and generalizability.
- Data Curation: Experimental ΔH<sup>ROP</sup> data was extracted from scientific literature using a natural language processing (NLP) pipeline.
- Computational Data Generation: A previously established first-principles computational scheme was used to generate ΔH<sup>ROP</sup> data. This method involves molecular dynamics simulations to explore diverse atomic configurations followed by ab initio calculations.
- Multitask Learning: A multitask ML model was developed to simultaneously learn from both experimental (ΔH<sup>ROP</sup><sub>expt</sub>) and computed (ΔH<sup>ROP</sup><sub>comp</sub>) data. This approach aims to transfer knowledge between the two data sources.
- Fingerprinting: Polymer structures were converted into machine-readable numerical representations (fingerprints) using a hierarchical fingerprinting procedure.
- Model Selection and Validation: Several ML algorithms (SVM, RF, BRF, GPR) were tested using a leave-one-out cross-validation (LOOCV) protocol. Gaussian Process Regression (GPR) was ultimately chosen for the production model due to its superior performance and ability to provide prediction uncertainty estimates.
Datasets
- Experimental Data: 109 ROP polymers with experimentally measured ΔH<sup>ROP</sup> values.
- Computed Data: ΔH<sup>ROP</sup><sub>comp</sub> data generated using first-principles calculations for a larger set of ROP polymers.
- Combined Dataset: A dataset comprising 193 unique ROP polymers with either experimental or computed ΔH<sup>ROP</sup> values, or both. This combined dataset was used for training the multitask ML model.
Key Insights and Ideas
- Data Fusion is Key: Combining limited but high-fidelity experimental data with more abundant but approximate computational data significantly improves the accuracy and generalizability of ML models for predicting polymer properties.
- Multitask Learning Advantage: By training on both experimental and computed data simultaneously, the model learns underlying correlations that enhance its ability to predict properties for unseen chemistries, especially when experimental data is scarce.
- Computational Data as a Scalable Pathway: While computationally expensive, first-principles calculations provide a scalable route to generate diverse data that can systematically expand the chemical space explored by ML models, leading to better predictions.
- GPR for Uncertainty Quantification: Gaussian Process Regression not only provides accurate predictions but also offers an estimate of prediction uncertainty, which is crucial for identifying regions where new computations or experiments are most needed.
Key Results
- The developed multitask ML model, based on GPR, achieved a test RMSE of 8.2 kJ/mol, approaching the "chemical accuracy" of ~5 kJ/mol.
- The model can predict ΔH<sup>ROP</sup> in essentially no time, a significant speedup compared to traditional computational methods.
- The accuracy of the model improved substantially with the inclusion of computed data, reducing RMSE from 12.2 kJ/mol (experimental data only) to 8.0 kJ/mol for the best multitask case (MT3).
- The NLP-based data extraction method increased the experimental ΔH<sup>ROP</sup> dataset by approximately 24%.
Limitations and Future Work
- Overfitting: The authors acknowledge that overfitting is still present, which is common with datasets of a few hundred data points.
- Data Scarcity: Performance is limited by data scarcity, as evidenced by the difference between test and train RMSEs and the lack of leveling in learning curves.
- Future Work: The authors plan to continuously update the production model by incorporating new computed ΔH<sup>ROP</sup> data. They also aim to extend the model to handle a wider range of chemistries and inform future experimental and computational efforts for designing depolymerizable and functional polymers.
Summary updated