Cross-sector deep learning scales life cycle assessment using unified textual descriptions

Life cycle assessment (LCA) is a foundational framework for quantifying global environmental impacts to guide decarbonization and circular economy transitions. Yet, conventional inventory compilation remains severely bottlenecked by data scarcity and intensive manual curation, with existing machine learning solutions typically confined to isolated industrial sectors rather than offering a scalable way to overcome these limitations. This fragmentation prevents pan-industrial scaling, thereby masking hidden environmental risks and hindering rapid, data-driven sustainability interventions. Here we develop LCA-TextNet, a generalizable deep learning framework that bypasses domain-specific feature engineering by predicting 25 life cycle impact indicators across 20 distinct industrial sectors directly from knowledge-based textual descriptions. Leveraging a Transformer-based architecture trained on over 16,000 datasets, the model maps high-dimensional text embeddings into uniform semantic spaces, achieving high accuracy (R² > 0.8) across 70% of sectors. Crucially, integrating application-domain stratification with an incremental learning strategy successfully mitigates distribution shifts during cross-version database validation, slashing climate change mean absolute error by 70%, from 2.0 to 0.6 kg CO₂-equivalent per unit. By exploiting the widespread asymmetry between ubiquitous descriptive text and scarce inventory metrics, this framework establishes a scalable paradigm that transforms textual knowledge into rapid, pan-industrial environmental intelligence to accelerate green transitions.