Modeling of coating ingredients via group contribution methods based on machine learning

Abstract

This thesis develops a data-driven framework for predicting polymer properties using machine learning and chemically informed molecular representations. Polymer performance arises from complex interactions between molecular structure and formulation composition, which makes property prediction challenging. It can be seen as a significant improvement compared to often applied empirical trial-and-error approaches in materials development. To address this challenge, this work integrates curated experimental datasets, graph-based molecular learning, and interpretable machine learning techniques to model structure-property relationships in polymer systems. UV coatings are used as a representative case study to investigate how molecular topology and formulation composition influence macroscopic material properties. Two complementary modeling approaches were adopted. First, the StructMon dataset was employed alongside graph neural networks, specifically the AttentiveFP architecture, to investigate the influence of monomer molecular structure on coating performance. Second, the CompMix dataset was analyzed using automated machine learning to explore the effect of compositional variation and monomer ratios. To enhance scientific transparency, interpretability strategies were applied to extract mechanistic insights and quantify the contributions of individual components. The results show that intrinsic properties governed by material properties, such as glass transition temperature and mechanical modulus, can be predicted with high accuracy when high-fidelity structural data is available. In contrast, extrinsic properties such as gloss proved more difficult to model because they depend on processing and environmental factors. Attention-weight analysis revealed that models internalized chemically meaningful substructures, whereas the inclusion of theoretical priors, such as the Fox equation, failed to improve performance, demonstrating also the necessity to apply the appropriate data. This work demonstrates that high-fidelity data and chemically aware representations are prerequisites for successful modeling and positions interpretability as a key enabler of rational formulation design in polymer science.