AI model turns text into life cycle impact estimates

Researchers from Tsinghua University, The University of Hong Kong and industry partners developed LCA-TextNet, a deep learning model that estimates 25 environmental impact indicators from text across 20 sectors. The system could speed life cycle assessment by filling data gaps, screening early-stage designs and flagging when predictions need expert review.

Why it matters: - Life cycle assessment is the standard way to measure the environmental footprint of products and processes, but compiling the data is slow, expensive and often incomplete. - A text-based model could help companies and researchers estimate impacts faster when inventory data are missing or scattered across documents. - The approach could make screening-level assessments more practical for early design work, policy analysis and ESG reporting.

What happened: - Researchers from Tsinghua University, Shanghai E-Carbon Digital Technology Co., Ltd., Shanghai HiQ Smart Data Co., Ltd. and The University of Hong Kong developed LCA-TextNet. - The model predicts life cycle impact assessment results from knowledge-based text descriptions. - The paper was accepted June 16, 2026, in Environmental Science and Ecotechnology. - The source article identifies Professor Shanying Hu and Zhijun Gui as the project leaders, with Kai Zhao and Biao Luo as first authors.

The details: - LCA-TextNet is a Transformer-based model trained on more than 16,000 activity datasets from ecoinvent version 3.10. - The model uses seven categories of textual information as input. - The system maps text embeddings into a shared semantic space instead of relying on sector-specific handcrafted features. - LCA-TextNet reached R² above 0.8 in 70% of sectors and for 17 of 25 environmental impact indicators. - The strongest results came from data-rich sectors with more consistent language, including waste treatment and recycling, and wood products. - Performance was weaker in transport, water supply and land use because of smaller sample sizes and more varied descriptions. - The team added an applicability-domain check to flag out-of-distribution predictions and help users decide when expert review is needed. - On newly introduced ecoinvent version 3.12 data, incremental learning cut climate change mean absolute error by 70%, from 2.0 to 0.6 kg CO₂ equivalent per unit. - The framework can also help complete background databases by filling gaps where inventory items lack matching entries. - When inventory data are unavailable, the model can estimate impacts directly from functional-unit descriptions. - The code is available on the HiQLCD GitHub repository. - Researchers need valid LCA database licenses to retrain and apply the framework within license limits. - The work was supported by the National Natural Science Foundation of China, grant No. U24B6016.

Between the lines: - The main advance is not replacing conventional LCA, but making the first pass faster and less dependent on manual feature engineering. - The results suggest text is a useful proxy for structured inventory data when the underlying sector has enough examples and consistent terminology. - The weaker sectors show the limits of the approach when language is sparse, noisy or highly heterogeneous. - The applicability-domain layer is important because it turns the model into a screening tool rather than an unchecked black box.

What's next: - The model could be used to support early-stage product design, policy studies and ESG workflows where detailed inventory data are not yet available. - Wider use will depend on access to licensed LCA databases and how well the framework transfers to new sectors and datasets. - The incremental-learning result suggests the model can improve as new inventory data arrive.

The bottom line: - LCA-TextNet turns descriptive text into a faster route to environmental impact estimates, with the strongest value in filling data gaps and accelerating preliminary assessments.