Utilizing machine learning to accelerate the identification and quantification of plastics or microplastics via only their basic elemental compositions

Accurate identification of polymer types and quantification of their contents for mixture samples are critical for recovering waste plastics and mitigating the environmental and human health risks posed by microplastics. However, it remains a major challenge due to the limitations of conventional spectroscopic and imaging techniques, which are costly and time-consuming. Here, a machine learning framework was established to quantify the contents of most commonly seen plastics present in mixed plastic samples using only elemental information, namely the contents of carbon (C), hydrogen (H), oxygen (O), and nitrogen (N). A synthetic dataset encompassing the major polymers was constructed based on theoretical elemental composition data. Different machine learning algorithms were compared using elemental descriptors as input features. SHapley Additive exPlanations analysis based on an optimal random forest model (for predicting polyethylene terephthalate, polyethylene & polypropylene, polyvinyl chloride, polystyrene, polyamide, and polycarbonate, average test R of 0.98) indicated that H and atomic ratios (H/C and O/C) were the most discriminative features, reflecting intrinsic chemical differences among polymers. Experimental validation using real mixed-plastic samples yielded an overall R of ∼0.70, and leave-one-out cross-validation performed solely on experimentally measured samples gave comparable performance. After further incorporating poly(butylene terephthalate), polyoxymethylene, poly(methyl methacrylate), and polylactic acid, the extended-feature ten-component model achieved an average test R of 0.89. This work proposed a promising methodology to address the critical gap in the quantitative determination of polymer mass fractions in multi-component plastic mixtures.