Intelligent application of large language model to life cycle assessment methodology

Abstract

Traditional life cycle assessment (LCA) methods face limitations such as high dependence on manual labor and expertise, difficulties in retrieving background databases, and low efficiency in manually writing carbon footprint reports, which can be addressed by large language models (LLMs). However, current research focused on a limited number of specific areas. This paper proposed a ubiquitous method for LLMs enhanced LCA research including knowledge summarization, data retrieval, and report generation. Firstly, a professional question-answering (QA) tool with knowledge summarization in LCA was realized utilizing knowledge base and retrieval-augmented generation method to alleviate hallucinations. Then, a natural language retrieval method for querying life cycle inventory (LCI) databases was developed by enhancing the LLMs with few-shot Chain of Thought and Chain of Code in Text2SQL. Furthermore, with code-interpreter, the LLM agent was able to auto-generate carbon footprint reports following a template. Finally, multiple experiments were performed to evaluate the performance of the different methods. The QA performance was evaluated using a QA dataset, achieving the BERTScore of 0.85, which demonstrates its capability in knowledge acquisition and summarization. The ability of database retrieval was validated by execution accuracy, and the highest value 0.9692 was obtained after augmented prompt engineering. The accuracy and completeness of the report were assessed through five dimensions, with the proposed model performing best in four of them. This paper provides a theoretical basis and experimental verification for the intelligent transformation of the LCA method. Future research will focus on enhancing interpretability of LLMs, expanding LCA-specific datasets, and improving intelligent LCI management to advance sustainable development.