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

Area of Science:

• Environmental Science
• Data Science
• Machine Learning

Background:

• Life cycle assessment (LCA) is crucial for sustainability but hindered by data scarcity and manual effort.
• Current machine learning solutions lack scalability across diverse industrial sectors, limiting broad application.
• This fragmentation obscures environmental risks and slows data-driven sustainability efforts.

Purpose of the Study:

• To develop a generalizable deep learning framework, LCA-TextNet, for predicting life cycle impact indicators.
• To overcome data scarcity and manual curation bottlenecks in conventional LCA inventory compilation.
• To enable pan-industrial scaling of environmental intelligence for accelerated green transitions.

Main Methods:

• Developed LCA-TextNet, a Transformer-based deep learning framework.
• Trained the model on over 16,000 datasets, predicting 25 impact indicators across 20 industrial sectors.
• Integrated domain stratification and incremental learning to handle database version changes and mitigate distribution shifts.

Main Results:

• Achieved high accuracy (R² > 0.8) across 70% of sectors by mapping text embeddings to semantic spaces.
• Significantly reduced climate change mean absolute error by 70% (from 2.0 to 0.6 kg CO₂-eq/unit) during cross-version validation.
• Demonstrated the framework's ability to transform textual knowledge into environmental intelligence.

Conclusions:

• LCA-TextNet provides a scalable paradigm to overcome data limitations in LCA.
• The framework enables rapid, pan-industrial environmental intelligence, accelerating decarbonization and circular economy transitions.
• Leveraging ubiquitous text data transforms environmental assessment for data-driven sustainability interventions.