Using Multimodal Large Language Models for Extracting Structured Life-Cycle Inventory Data from Scientific Literature

Abstract Life-cycle inventory (LCI) construction remains labor-intensive because relevant information is distributed across heterogeneous text, tables, and figures and is reported using inconsistent terminology, units, and system boundaries. Although large language models (LLMs) offer opportunities for automated information extraction, LCI-specific workflows that reliably convert literature-reported information to structured and traceable inventory records remain limited. Here, we present LCIMiner, an automated LLM-based workflow integrating literature retrieval and screening, text-based selection of LCI-relevant figures and tables, multimodal information extraction, task-specific self-review, and structured LCI construction without model retraining. We evaluated LCIMiner in aquaculture, polymer production, and biofuel production, which differ substantially in their production processes, inventory structures, and reporting conventions. Across these domains, LCIMiner achieved higher end-to-end extraction performance than matched single-model baselines and converted heterogeneous literature evidence to machine-readable, source-traceable LCI records. We further applied a separate human-guided harmonization procedure to organize the extracted aquaculture records into a cross-study analytical data set while preserving the original source information. Together, the automated extraction workflow and subsequent human-guided harmonization provide a transparent and extensible approach for developing reusable foreground LCI data sets from heterogeneous scientific literature.