Utilizing machine learning to accelerate the identification and quantification of plastics or microplastics via only...

Area of Science:

• Polymer Science
• Analytical Chemistry
• Machine Learning

Background:

• Accurate polymer identification and quantification in mixtures are vital for waste plastic recycling and mitigating microplastic pollution.
• Conventional methods are often costly and time-consuming, hindering efficient analysis.

Purpose of the Study:

• To develop a machine learning framework for quantifying common plastic types in mixtures using only elemental composition.
• To establish a rapid and cost-effective method for analyzing complex plastic waste.

Main Methods:

• A synthetic dataset of major polymers was created based on theoretical elemental compositions (C, H, O, N).
• Various machine learning algorithms were trained and compared using elemental data as input features.
• SHapley Additive exPlanations (SHAP) identified hydrogen and atomic ratios (H/C, O/C) as key discriminative features.

Main Results:

• An optimal random forest model achieved an average R² of 0.98 for predicting six common polymers.
• Experimental validation on real mixed plastics yielded an R² of approximately 0.70.
• An extended ten-component model incorporating additional polymers reached an average R² of 0.89.

Conclusions:

• Elemental analysis combined with machine learning provides a promising, efficient approach for quantifying polymer mass fractions in multi-component mixtures.
• This methodology addresses a critical gap in the analysis of plastic waste, facilitating improved recycling and environmental management.