From coarse-grained molecular dynamics to machine learning: A comparative study of modeling techniques for predicting polycarbonate maximum stress from structural features

T Leelaprachakul and Y Umeno, COMPUTATIONAL MATERIALS SCIENCE, 257, 113971 (2025).

DOI: 10.1016/j.commatsci.2025.113971

This study explores the use of machine learning to predict the maximum stress of polycarbonate materials based on key structural parameters: polydispersity, radius of gyration, and molecular entanglements. These parameters are derived from coarse-grained molecular dynamics simulations of transverse constrained tension. We compare various machine learning algorithms, including multivariate linear regression, log-log regression, gradient boosted regression trees, extreme gradient boosted regression trees, and artificial neural networks. Model performance is evaluated using the coefficient of determination R2, root-mean-squared error, computational efficiency. While traditional linear and log-log models achieve R2 values of 0.77 and 0.83, respectively, boosted regression trees and artificial neural networks attain higher R2 values of approximately 0.90. However, artificial neural networks require over four orders of magnitude more computation compared to boosted regression tree models, which has R2 of 0.89. To interpret feature importance, employ Shapley additive explanations analysis on the extreme gradient boosted model. The radius of gyration is identified as the most influential parameter, highlighting its role in maintaining structural integrity during tensile deformation. A larger radius of gyration enhances load distribution along molecular chains, ensuring uniform stress transfer and improved mechanical performance.

Return to Publications page