Methodology: AI-Driven LCA Data Extraction

1. Abstract

The construction and real estate sectors are under increasing pressure to calculate the carbon footprint of buildings using Life Cycle Assessments (LCA). The primary source of this data is the Environmental Product Declaration (EPD) formatted according to the EN 15804 standard. This paper details the architecture and methodology behind the EPD Extractor, a cloud-based SaaS tool designed to automate the extraction of critical Global Warming Potential (GWP) metrics from unstructured PDF documents using Large Language Models (LLMs).

2. The Challenge of Unstructured EPDs

While EN 15804 standardizes the metrics that must be reported (such as A1-A3, C3, C4, and Module D), it does not standardize the visual layout of the PDF. Manufacturers present data in diverse tabular formats, varying page orientations, and embedded graphics. Traditional Optical Character Recognition (OCR) and regex-based scraping tools fail consistently when confronted with this structural variance.

3. Solution Architecture: Gemini 3.1 Pro

To overcome layout inconsistency, EPD Extractor utilizes the Google Gemini 3.1 Pro model. Unlike traditional scrapers, Gemini possesses multimodal spatial reasoning capabilities. The system pipeline operates as follows:

1. Ingestion: The user uploads a PDF (up to 20 MB). A secure middleware validates the MIME type and file size.
2. Prompt Engineering: A highly restrictive, zero-shot system prompt instructs the LLM to locate specific metrics (Product Name, Manufacturer, Density, GWP A1-A3, C3, C4, Module D) and enforce a strict JSON output schema.
3. Vision-Language Processing: The LLM processes the PDF in its entirety, contextually differentiating between metric units (e.g., kg CO2 eq. vs. MJ) and lifecycle stages, regardless of tabular distortion.
4. Sanitization: The backend strips any conversational markdown generated by the LLM, parses the JSON, and delivers standard structured data to the frontend UI.

4. Accuracy and The "Human-in-the-Loop" Mandate

While the spatial reasoning of Gemini 3.1 Pro yields a high extraction success rate, generative AI is susceptible to hallucinations or misalignments when encountering non-standard appendices or highly fragmented tables.

Therefore, EPD Extractor is designed explicitly as a productivity enhancer, not an autonomous decider. The system enforces a strict Human-in-the-Loop (HITL) policy. Engineers must use the extracted CSV data as a baseline for cross-referencing against the original PDF, reducing initial data entry time by an estimated 85% while maintaining absolute engineering accountability.