To what extent can Large Language Models support Life Cycle Assessment of manufacturing processes?

Abstract

The industrial sector is responsible for approximately one quarter of global CO2 emissions each year, driving the scientific community to adopt quantitative tools such as Life Cycle Assessment (LCA) to evaluate environmental performance. However, traditional LCA is highly dependent on expert knowledge and time-intensive, especially for Life Cycle Inventory (LCI), which remains a major challenge, as process-specific information is often limited or based on estimations. This is critical in discrete manufacturing sector, which is highly variable and largely unstandardized. The rapid spread of Artificial Intelligence (AI) in the most diverse areas has led scholars to question whether such LCA challenges in manufacturing processes can be addressed by integrating Large Language Models (LLMs) into LCAs. This study investigates the potential of LLMs to support LCA development by evaluating their accuracy, reliability, and methodological consistency. Three approaches with different levels of LLM involvement were compared: (i) an expert knowledge-based LCA, (ii) an LLM-driven LCA, and (iii) an LLM-supported LCI. Prompt engineering techniques were applied to optimize LLM responses and reduce error propagation. Results showed that the LLM-supported LCI approach produced more accurate and consistent outcomes than the LLM-driven one, confirming that expert supervision is essential to ensure methodological rigor and data reliability. Nonetheless, limitations related to data variability, reliance on open-access sources, and the inability to handle multi-category indicators were identified. The study concludes that LLMs can effectively support early-stage LCA modeling and streamline data collection but cannot replace expert judgment and standard-compliant analytical procedures, especially within manufacturing processes context.