Identification of environmental microplastics using large language models: DeepSeek-R1-Distill-Llama-8B, GPT-4o, and GPT-4o-mini

Abstract

Microplastic pollution in the environment poses increasing risks to both ecological and human health. Identifying microplastics in environmental samples is important for monitoring and mitigation. However, current methods rely on manual interpretation of infrared (IR) spectra, which is time-consuming and labor-intensive. Thus, this study investigates the potential of large language models (LLMs) for identifying microplastics using IR spectra from environmental samples. Three models, DeepSeek-R1-Distill-Llama-8B, GPT-4o-2024–08–06 (GPT-4o), and GPT-4o-mini-2024–07–18 (GPT-4o-mini), were evaluated within a structured workflow that integrates spectral processing and model implementation. A performance evaluation framework was developed to measure identification accuracy. Results indicate that DeepSeek-R1-Distill-Llama-8B outperformed others, achieving an accuracy exceeding 0.93 across all tested polymer types, making it the preferred choice. GPT-4o proved a strong alternative, particularly when local execution is impractical, with accuracy above 0.86. GPT-4o-mini underperformed and is not recommended. Despite these promising outcomes, challenges persist, including the need to optimize spectral processing parameters and refine prompt design. As the first study to apply LLMs to microplastic identification, this work offers a foundational reference for leveraging LLM-driven spectral analysis in environmental monitoring.