What question did this study set out to answer?

The aim is to evaluate machine learning techniques for predicting the glass transition temperature of polymers based on structural attributes.

March 21, 2026Open Access

Comparative evaluation of machine Learning-based QSPR techniques for predicting polymer glass transition temperature

Key Points

The aim is to evaluate machine learning techniques for predicting the glass transition temperature of polymers based on structural attributes.
Used a dataset of 250 polymers to develop ML-QSPR models.
Employed Genetic Algorithm and Multiple Linear Regression to optimize molecular descriptors.
Conducted comparative analysis with non-linear methods like SVR, GPR, RF, and MLP.
MLP model achieved a training R² of 0.82 and external validation R² of 0.79.
The approach demonstrated that specific structural descriptors significantly enhance prediction accuracy.

Abstract

The glass transition temperature (\: T₆) is a pivotal parameter for amorphous polymers, influencing their behavior under different thermal properties. Assessing \: T₆ is crucial for evaluating material performance across varying temperatures. Our research integrates machine learning with cheminformatics framework to analyze \: T₆ values across a broad polymer dataset. This framework improves a comprehensive understanding of the quantitative correlations between the polymers’ structural attributes and their \: T₆. Utilizing a dataset of 250 polymers, we constructed a series of Machine Learning-based Quantitative Structure-Property Relationship (ML-QSPR) models. The initial ML-QSPR used Genetic Algorithm (GA) and Multiple Linear Regression (MLR) methods to identify an optimal set of molecular descriptors. Subsequently, various non-linear machine learning methods, including Support Vector Regression (SVR), Gaussian Process Regression (GPR), Random Forest (RF), and Multi-Layer Perceptron (MLP), were used for comparative purposes and predictive analysis. The results demonstrate that MLP, driven by seventeen (17) selected descriptors, yields the most accurate ML model, with training and external validation \: R^2\: values of 0. 82 and 0. 79, respectively. This model predicts with good performance the \: T₆ values of the polymers under study, showing the role of specific structural descriptors in refining polymer property predictions.

Perguntar à IA

Bookmark

View Full Paper