[Paper Review] AutoPCF: Efficient Product Carbon Footprint Accounting with Large Language Models

Zhu Deng, Jinjie Liu+6 more|arXiv (Cornell University)|Aug 8, 2023

Environmental Impact and Sustainability4 citations

TL;DR

This paper proposes AutoPCF, an AI-driven framework that leverages large language models (LLMs) to automate product carbon footprint (PCF) accounting for 'cradle-to-gate' life cycles. By using LLMs to generate life cycle inventories and matching emission factors via a semantics-based model, AutoPCF reduces modeling time from days to minutes, achieving a mean estimation error of 42% compared to expert benchmarks—demonstrating strong potential for scalable, efficient PCF assessment.

ABSTRACT

The product carbon footprint (PCF) is crucial for decarbonizing the supply chain, as it measures the direct and indirect greenhouse gas emissions caused by all activities during the product's life cycle. However, PCF accounting often requires expert knowledge and significant time to construct life cycle models. In this study, we test and compare the emergent ability of five large language models (LLMs) in modeling the 'cradle-to-gate' life cycles of products and generating the inventory data of inputs and outputs, revealing their limitations as a generalized PCF knowledge database. By utilizing LLMs, we propose an automatic AI-driven PCF accounting framework, called AutoPCF, which also applies deep learning algorithms to automatically match calculation parameters, and ultimately calculate the PCF. The results of estimating the carbon footprint for three case products using the AutoPCF framework demonstrate its potential in achieving automatic modeling and estimation of PCF with a large reduction in modeling time from days to minutes.

Motivation & Objective

• To address the time-consuming and expertise-intensive nature of traditional life cycle assessment (LCA) for product carbon footprint (PCF) estimation.
• To investigate whether large language models (LLMs) can emergently model product life cycles and generate accurate input-output inventories for PCF calculation.
• To develop and evaluate an automated AI-driven framework, AutoPCF, that integrates LLMs with emission factor matching to accelerate and standardize PCF estimation.
• To assess the reliability and accuracy of LLM-generated PCF estimates compared to expert-based models across diverse industrial products.

Proposed method

• Employed five general-purpose LLMs (including GPT-3.5, GPT-4, and Tongyi Qianwen) to automatically generate 'cradle-to-gate' life cycle processes, inputs, and outputs for three industrial products.
• Applied a semantics-based matching model to map raw materials and energy inputs to corresponding emission factors from databases.
• Integrated deep learning-based parameter matching to link activity data (e.g., material quantities) with appropriate emission factors for CO2-eq calculation.
• Used two activity data generation approaches (DGA and manual input) to evaluate model robustness and consistency.
• Calculated total PCF by summing CO2-eq emissions from all inputs and processes, using standardized conversion factors.
• Validated results against expert-generated PCF estimates to assess accuracy and error margins.

Experimental results

Research questions

• RQ1 Can large language models (LLMs) effectively model the 'cradle-to-gate' life cycle processes of industrial products with minimal human input?
• RQ2 How accurate are LLM-generated life cycle inventories (LCI) compared to expert-constructed inventories in terms of input-output structure and emission factor matching?
• RQ3 What is the performance of different LLMs in generating activity data and matching emission factors for PCF estimation?
• RQ4 How does the AutoPCF framework reduce modeling time and effort compared to traditional expert-based LCA?
• RQ5 To what extent do uncertainties in LLM outputs affect the reliability of PCF estimates, and how can they be mitigated?

Key findings

• The AutoPCF_GPT-3.5_DGA model achieved the most stable and accurate results across all three products, with a mean estimation error of 42% compared to expert benchmarks.
• GPT-3.5, GPT-4, and Tongyi Qianwen demonstrated emergent capabilities in modeling production processes (F1-score > 0.4) and activity inventories (F1-score > 0.35), indicating strong potential for PCF automation.
• The framework reduced PCF modeling time from days to minutes, significantly accelerating the estimation process.
• Expert models showed high variability in results (CV: 24–58%), highlighting the need for standardized, automated tools like AutoPCF to improve consistency.
• LLMs exhibited limitations in generalization and consistency, particularly in emission factor matching and activity data precision, underscoring the need for improved training data and prompt engineering.
• The study confirms that LLMs can serve as a viable, scalable alternative to expert-driven LCA when combined with structured matching and validation mechanisms.