Assessing Uncertainty in Machine Learning for Polymer Property Prediction: A Benchmark Study.

Area of Science:

• Materials Science
• Computational Chemistry
• Data Science

Background:

• Machine learning (ML) accelerates materials discovery but requires reliable predictions.
• Uncertainty quantification (UQ) is vital for trustworthy ML in high-stakes applications like functional polymer design.
• Evaluating UQ methods is crucial for optimizing ML model performance and reducing experimental costs.

Purpose of the Study:

• To comprehensively evaluate nine UQ methods for ML in polymer property prediction.
• To assess UQ method performance across diverse datasets, including out-of-distribution (OOD) and specific polymer types.
• To provide guidance on selecting optimal UQ strategies for accelerating functional polymer discovery.

Main Methods:

• Evaluated nine UQ methods: Ensemble, GPR, MCD, MVE, BNN-VI, BNN-MCMC, EDL, QR, and NGBoost.
• Predicted key polymer properties: glass transition temperature (Tg), band gap (Eg), melting temperature (Tm), and decomposition temperature (Td).
• Assessed models using R², Spearman's rank correlation, and calibration area across various data scenarios.

Main Results:

• Optimal UQ method selection is context-dependent.
• Ensemble method excelled in general in-distribution predictions.
• BNN-MCMC showed strong performance for OOD scenarios; NGBoost was top for high-Tg polymers.
• BNN-VI demonstrated consistent superiority across nine polymer classes.

Conclusions:

• Tailored UQ strategies are critical for enhancing ML prediction trustworthiness in polymer science.
• Effective UQ selection optimizes experimental validation and accelerates the discovery of advanced functional polymers.
• This benchmark provides a robust framework for UQ method selection in materials informatics.