Smarter LCA : Using large language models to predict manufacturing processes and their uncertainties☆

Prospective Life Cycle Assessment offers significant potential to guide sustainable product development by influencing design in the early stages, when the flexibility for impactful changes is greatest. However, its practical application is often constrained by the limited availability and quality of foreground life cycle inventory data. For mechanically engineered products, key manufacturing-related decisions that determine material waste and energy demand are frequently unavailable, and while expert estimations can fill these gaps, this approach is time-consuming and limited by specialist availability. To overcome this challenge, we present an automated, uncertainty-aware method for generating screening-level missing manufacturing foreground data. Our approach leverages large language models, starting with a fine-tuned Bidirectional Encoder Representations from Transformers model that predicts part group classifications directly from short part names. These predicted groups are then mapped to specific manufacturing process datasets in ecoinvent 3.9.1 using a deterministic categorical mapping rule, with the large language model used to extract manufacturing parameters from process descriptions. A key feature of our method is the automated generation of pedigree matrix scores to quantify prediction confidence and mapping quality. This uncertainty is then propagated using Monte Carlo simulation, providing transparent decision support. An illustrative industrial case study shows that the pipeline can predict plausible proxies for manufacturing processes from limited textual data, enabling a more robust and timelier prospective Life Cycle Assessment when crucial manufacturing information remains unknown.