Predicting glass transition temperature of polymers by combining molecular dynamics simulations and machine learning techniques

Abstract

Polymeric materials, valued for their cost-effectiveness and ease of processing, play a pivotal role in diverse applications. Among them, solution polymerized styrene butadiene rubber (SSBR) stands out with low rolling and wet skid resistance, ideal for eco-friendly tire. The glass transition temperature (Tg) is an important parameter in determining the performance of SSBR, which is time-consuming to test with conventional experiments. In recent years, machine learning (ML) has emerged as a useful tool in materials science. In this study, all-atom molecular dynamics (AAMD) simulation provided the foundational training data. Subsequently, ML techniques are utilized to efficiently predict SSBR’s Tg based on its content of four key structural components. To address the challenge of limited sample sizes, the innovative integration of Ordinary Kriging (OK) from geostatistics and Nearest Neighbor Interpolation (NNI) was employed to compute label values for interpolation points, which surpasses previous studies relying solely on NNI. Comparative analysis of four ML models (XGBoost, Ridge, LASSO, and SVR) reveals XGBoost’s superior performance in handling this nonlinear problem, with consistently high R2 and low RMSE scores. Through 900 repeated experiments, the model demonstrates robustness, maintaining an average R2 of approximately 0.9703. Ultimately, the novel integrated ML approach (NNI-OK-XGBoost) was proposed. Furthermore, four new combinations of structural unit content were evaluated using both ML predictions and AAMD simulations. The validation confirms the efficiency and accuracy of this method in predicting SSBR’s Tg across varying composition ratios within a specified composition space. This work presents a viable solution for accurately and rapidly predicting the Tg of polymeric materials, particularly in scenarios with limited sample sizes.