Comparison of Machine Learning Methods towards Developing Interpretable Polyamide Property Prediction.

Area of Science:

• Polymer Science
• Materials Science
• Computational Chemistry

Background:

• Polyamides possess desirable thermal, mechanical, and chemical properties, but their diverse structures complicate traditional quantitative structure-property relationship (QSPR) analysis.
• Predicting properties of polyamides with varying structures, from linear to aromatic, is challenging using conventional methods.

Purpose of the Study:

• To develop high-fidelity machine learning models for predicting key polyamide properties: glass transition temperature (Tg), melting temperature (Tm), density (ρ), and tensile modulus (E).
• To evaluate the effectiveness of different fingerprints, including extended connectivity fingerprints (ECFP), and machine learning algorithms for QSPR modeling of polyamides.

Main Methods:

• Employed machine learning algorithms, including random forest and linear regression, with various QSPR fingerprints (ECFP and traditional).
• Utilized feature selection and regularization techniques to enhance the performance of linear models.
• Assessed prediction accuracy for Tg, Tm, ρ, and E across different modeling approaches.

Main Results:

• Non-linear random forest models generally outperformed linear regression.
• Linear models achieved comparable accuracy to non-linear models when enhanced with feature selection or regularization.
• Accurate prediction of tensile modulus (E) was not achieved, potentially due to data heterogeneity and measurement challenges.
• QSPR models identified the fraction of rotatable bonds and rotational degree of freedom as key factors influencing polyamide properties.

Conclusions:

• Machine learning, particularly QSPR models, can effectively predict key polyamide properties (Tg, Tm, ρ), aiding material design.
• While complex models show high accuracy, simpler QSPR models provide valuable chemical intuition for property modification.
• Further research is needed to address challenges in predicting tensile modulus for polyamides.