Uncertainty Quantification in Machine Learning for Glass Transition Temperature Prediction of Polymers

Machine learning (ML) has emerged as a vital tool in polymer informatics, accelerating the discovery and design of novel polymers while reducing experimental costs [1]. Uncertainty quantification (UQ) is essential for accurate predictions and innovative polymer design through ML [2]. In this study, we evaluate six UQ methods: ensemble, Gaussian process regression (GPR), Monte Carlo dropout (MCD), mean-variance estimation (MVE), Bayesian neural network (BNN), and evidential deep learning (EDL) for predicting the glass transition temperature of homopolymers. We assess accuracy and performance using R², Spearman’s rank correlation, and calibration metrics, providing a comprehensive UQ evaluation. Our evaluation includes both test data, consisting of polymer samples with known glass transition temperatures used for model validation, and out-of-distribution (OOD) data, derived from experimental results and molecular dynamics simulations, representing polymer structures outside the training set's chemical space. Among the six UQ methods, BNN performs best due to its strong accuracy, reliable error-uncertainty relationship, and well-calibrated predictions, making it a promising tool for polymer design. This study lays the groundwork for developing UQ models that, when integrated with active learning, can iteratively guide the discovery and optimization of novel polymers, reducing the need for costly and time-consuming experiments.