Co-pyrolysis of sludge and plastics has gradually emerged as a crucial technical approach for waste reduction and resource recovery. This study develops high-precision, interpretable prediction models and quantifies the contributions of core risk factors to environmental risks. Based on the experimental datasets from 2015 to 2025, which include operational parameters and eight potential toxic elements (PTEs) with four chemical speciation fractions: acid-soluble/exchangeable (F1), reducible (F2), oxidizable (F3), and residual (F4), we constructed six machine learning models. Based on the experimental datasets from 2015 to 2025, which include operational parameters and eight potential toxic elements (PTEs) chemical speciation (F1–F4), we constructed six machine learning models. Feature importance analysis and Shapley Additive Explanation (SHAP) analysis were employed to identify core risk factors and interpret the model’s decision logic. Results indicate that XGBoost, Random Forest and CatBoost outperform other models, achieving test accuracies of 0.94, 0.92, and 0.90, with weighted F1-Scores of 0.94, 0.92, and 0.90, respectively. Feature importance highlights the most important features for the six different models, with Cd-F4, As-F1, and Cu-F4 contributing most significantly to the model predictions. SHAP analysis quantified the contributions of each feature to the model predictions, verified Cd-F4 as the primary risk discriminant, and further revealed that F1 and F4 of PTEs are key factors in distinguishing risk levels. This study proposes an interpretable machine learning framework, providing a theoretical basis for the optimization of the sludge and plastic co-pyrolysis process and the assessment of potential risks.

Continuous cropping in tropical protected cultivation can impair soil health and increase the risk of soil-borne disease. We conducted a single-season field comparison in Hainan (October 2025–February 2026) to characterize the performance and rhizosphere microbial profiles of the mini-watermelon cultivar ‘Meiyue’ grown either on its own roots or grafted onto Cucurbita ficifolia. Plant growth, leaf gas exchange, fruit quality, and the field incidence of Fusarium wilt-like symptoms were recorded, and bacterial 16S V3-V4 and fungal ITS2 amplicons from the rhizosphere were sequenced. Grafted plants were taller and had thicker stems and longer internodes. Leaf chlorophyll content, stomatal conductance, and intercellular CO₂ concentration were higher in grafted plants, whereas net photosynthetic rate did not differ between treatments. Wilt-like symptoms occurred in 55.41% of own-rooted plants, while no grafted plants developed symptoms during the trial. Mean single-fruit weight was 2.8-fold higher in the grafted treatment; grafted fruits were also larger and had a higher central soluble solids content. Bacterial and fungal beta diversity did not differ significantly between treatments (PERMANOVA, p = 0.100), although several taxa differed in relative abundance and LEfSe identified treatment-associated biomarkers. An exploratory functional-profile sensitivity analysis, performed after excluding one high-leverage own-rooted replicate, indicated modest differences in predicted nitrogen-transformation functions and fungal trophic modes; these predictions were not subjected to inferential testing. Correlations between microbial relative abundance and fruit traits were interpreted as associations rather than evidence of function. Within the constraints of a single-season, two-treatment study, watermelon grafted onto C. ficifolia showed better field performance and treatment-associated shifts in selected rhizosphere taxa under continuous-cropping conditions.