A scalable, dispersion-aware framework for quantifying clustering severity in polymer nanocomposites and assessing mechanical impact

Abstract

Identifying a reliable threshold for cluster size in polymer nanocomposites remains a key challenge, limiting quantitative evaluation of dispersion and its effect on mechanical behavior. To address this, a novel Clustering Propensity Index (CPI) was introduced for automated quantification of nanoparticle clustering in polyoxymethylene (POM)/carbon black (CB)/calcium carbonate (CC) nanocomposites. Deep learning-based segmentation using YOLOv8 was first applied to SEM micrographs to extract particle area distributions (PADs). CPI was then computed by applying kernel density estimation (KDE) with adaptive bandwidths to define optimal bin intervals. As an innovative approach, a weighted frequency analysis, emphasizing larger particles, was used to quantify clustering severity. According to the results, CPI values dropped significantly to 0.0058 and 0.0040 at 1.5 and 3 wt% CC-representing 72.90 % and 81.31 % reductions from POM/CB (0.0214)-but increased to 0.0403 at 4.5 wt% due to re-agglomeration. An XGBoost-based two-variable model with CC content and CPI, as the inputs, was employed to predict mechanical responses. Feature importance analysis revealed CPI2 as the most influential factor for impact toughness (SHAP approximate to 0.4), while CC content governed stiffness. The proposed framework provides a scalable, dispersion-aware methodology for quantifying clustering and systematically assessing its impact on mechanical properties.