Multimodal large language models for big data synthesis: Addressing the data bottleneck of plastic chemicals

The global plastic pollution is escalating, driven in part by hazardous plastic chemicals. However, the absence of comprehensive data on chemical exposure and hazards, particularly concerning economic indicators, physicochemical properties, environmental behavior parameters, toxicity profiles, and regulation status, severely constrains effective decision-making and policy formulation. This study developed a multimodal large language model-based framework for extracting economic indicators of plastic chemicals from unstructured sources. From 36,289 images and 969 literature sources, 44,249 entries covering 814 chemicals were extracted across 9 economic indicators, including production volume, production value, production capacity, market size, consumption, and import/export volumes and values. The pipeline achieved an average F₁ score of 92.8% (95% confidence interval: 89.9–95.6%), demonstrating high accuracy and comprehensiveness. Toxicity data for 4,551 chemicals, environmental behavior parameters for 9,770 chemicals, and regulatory information for 13,684 chemicals from 111 global lists were also manually compiled. A graph attention network model was employed to predict 11 physicochemical properties as well as the persistence, bioaccumulation, mobility, and toxicity (PBT/PMT) properties of 9,768 chemicals. This study constructed the most up-to-date and comprehensive dataset with multiple data types for plastic chemicals, providing an AI-based methodological foundation for synthesizing big data to support the sound management of plastic chemicals.