Machine learning-driven estimation of microplastic percentage yield for rapid and accurate quantification

Abstract

Accurate microplastic (MP) quantification in agricultural soils is critical for environmental risk assessment, yet variability in extraction efficiency remains a significant barrier. This study This study investigated machine learning (ML) algorithms to predict MP percentage yield and develop interval-wise correction factors for improved MP quantification. Soils were spiked with four MP types and subjected to density separation using seven brine solutions (e.g., zinc chloride and sodium iodide; 1.00–1.58 g/cm3). Pearson correlation and feature importance scores identified brine density as the dominant driver (score = 0.52), while leave-one-condition-out (LOCO) analysis further validated the robust non-linear superiority of ensemble ML models. Among the six ML model, the random-forest (RF) algorithm exhibited the highest coefficient of determination (R2) at 0.991 with the lowest error metrics. Furthermore, RF-driven interval-wise correction factors demonstrated superior adjustment efficiency, effectively aligning predicted yields with target values across diverse recovery scenarios. The proposed integrated framework offers a scalable approach for supporting standardized extraction protocols. The findings of this study can help for an understanding of the complex interaction between experimental parameters and recovery MP yields, ultimately facilitating more precise laboratory-scale MP monitoring.