AI-Driven engineering product data structuring for LCA and foreground LCI modeling

Life Cycle Assessment (LCA) is widely used to evaluate the environmental performance of industrial products, yet the generation of the Life Cycle Inventory (LCI), the phase in which environmental impact ultimately originates, remains largely manual. Engineering documents such as Bills of Materials (BoMs) contain detailed information about product composition and hierarchy, but this information is created for manufacturing purposes, is structurally heterogeneous across companies and even across products from the same manufacturer, and encodes product hierarchy only implicitly, typically through a numeric Level attribute and row ordering. Interpreting this implicit structure before physical disassembly and material identification can proceed is a recurring, expert-dependent source of manual effort, as documented through the teams's own industrial practice in the household appliance sector. This research work proposes and validates an AI-driven methodology that automatically transforms heterogeneous industrial BoMs into structured, hierarchical, machine-readable representations suitable for foreground LCI development. The methodology combines a deterministic preprocessing stage, which detects the BoM header and metadata regions and dynamically generates a document-specific schema from the attributes actually present, with a Large Language Model (LLM)-based stage that extracts component information and reconstructs parent–child relationships from the BoM's Level attribute according to explicit, prompt-embedded hierarchy rules. The methodology is evaluated through a technical validation across six BoMs of varying structure and scale, combining qualitative stage-level checks with quantitative Precision, Recall, and F1-score measures for both attribute extraction and hierarchical reconstruction, and through an industrial case study on a combination oven–microwave, comparing the automated output against a reference inventory obtained by manual disassembly. Results show that attribute-level extraction is consistently reliable across formats, while hierarchy reconstruction is strong under moderate scale and an explicit numeric signal but degrades under two identified conditions: large, deeply nested BoMs and BoMs lacking a hierarchy indicator altogether. Within the industrial case study, the automated output fully reproduced the verified structure of three product modules spanning up to five hierarchy levels. These findings demonstrate that AI can meaningfully reduce the manual interpretation effort preceding foreground LCI preparation, while clarifying the specific structural conditions under which its reliability currently holds and where expert oversight remains necessary.