Degradable and non-degradable microplastics regulate soil greenhouse gas emissions: Multi-factor insights and key monitoring factors from a Kolmogorov-Arnold Network ensemble model.

Microplastics (MPs) are increasingly recognized as a potential factor influencing soil greenhouse gases (GHGs) emissions, yet their effects remain inconsistent due to complex interactions among soil properties, microbial communities, and MP characteristics. To address this, a dataset including CO₂, N₂O, and CH₄ emissions was constructed, and multiple machine learning models (KNN, RF, SVM, BPNN, XGB, and KAN) along with their ensemble forms were applied to identify optimal predictive frameworks. On this basis, SHAP analysis, Spearman correlation heatmaps, feature clustering, and variance inflation factor (VIF) analysis were integrated to examine inter-variable relationships and quantify feature contributions under high-dimensional and multicollinear conditions. The results indicate that KAN and tree-based models effectively capture nonlinear relationships, while ensemble strategies generally improve model performance depending on data structure. GHG emissions were primarily controlled by environmental variables such as incubation time, soil water content (SWC), dissolved organic carbon (DOC), soil texture, and microbial groups. In contrast, MPs played a secondary and mainly indirect role, mediated through physicochemical and microbial pathways, with more evident effects observed for CH₄ than for CO₂ and N₂O. These findings highlight the context-dependent nature of MP impacts and offer a useful reference for further investigations into complex soil systems.