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

Area of Science:

• Environmental Science
• Analytical Chemistry
• Data Science

Background:

• Accurate microplastic quantification in agricultural soils is crucial for environmental risk assessment.
• Variability in extraction efficiency hinders reliable microplastic (MP) analysis.

Purpose of the Study:

• Investigate machine learning (ML) algorithms to predict MP percentage yield.
• Develop interval-wise correction factors for enhanced MP quantification in soils.

Main Methods:

• Spiked soil samples with four MP types and used density separation with seven brine solutions (1.00-1.58 g/cm³).
• Employed Pearson correlation, feature importance, and leave-one-condition-out (LOCO) analysis.
• Evaluated six ML models, focusing on the random-forest (RF) algorithm.

Main Results:

• Brine density was identified as the dominant factor influencing MP yield (feature importance score = 0.52).
• Ensemble ML models demonstrated non-linear superiority over traditional methods.
• The RF algorithm achieved a high coefficient of determination (R² = 0.991) with minimal error.
• RF-derived correction factors significantly improved the accuracy of predicted MP yields.

Conclusions:

• An integrated ML framework offers a scalable solution for standardized MP extraction protocols.
• This approach enhances understanding of parameter interactions and MP recovery, enabling precise laboratory monitoring.
• Improved MP quantification supports more robust environmental risk assessments in agricultural settings.