A glimpse inside materials: Polymer structure – Glass transition temperature relationship as observed by a trained artificial intelligence

Abstract

Artificial neural networks (ANNs), a subset of Quantitative Structure-Property Relationship (QSPR) methods, offer a promising avenue for addressing challenges in materials science. In particular, ANNs can learn intricated patterns within the experimental data, enabling them to predict properties and recognize complex relationships with remarkable accuracy. However, the opacity of ANNs, normally acting as black boxes, raises concerns about their reliability and interpretability. To enhance their transparency and to uncover the underlying relationships between chemical features and material properties, we propose a novel approach that employs Gradient-weighted Class Activation Mapping (Grad-CAM) applied to Convolutional Neural Networks (CNNs). By analyzing these attention maps, we identify the crucial chemical features influencing the prediction of a polymer property, specifically the glass transition temperature (Tg). Our methodology is validated using a dataset of atactic acrylates, allowing us to not only predict Tg values for a control group of polymers but also to quantitatively assess the impact of individual monomer structural elements on these predictions. This work proposes a step towards transparent models in materials science, contributing to a deeper understanding of the intricate relationship between chemical structures and material properties.