Accelerating life cycle assessment: A machine learning-based decision-support pipeline for multi-source environmental impact prediction

Life Cycle Assessment (LCA) is a well-established sustainability standard, yet its application is hindered by extensive data requirements and time-intensive inventory modeling. Machine Learning (ML) can accelerate impact prediction, but current LCA-ML integrations struggle with the structural heterogeneity of multi-source datasets. This study develops and validates a high-throughput decision-support pipeline that harmonizes 1219 compiled datasets from three LCA databases (Agribalyse, ecoinvent, and Agri-footprint) into 913 training samples and predicts 11 midpoint impact categories. Thirteen ML algorithms were compared under repeated stratified cross-validation. Gradient Boosting architectures (CatBoost, XGBoost, and GBM) consistently outperformed the remaining models, frequently achieving coefficients of determination (R 2 ) above 0.80, and tolerated aggressive dimensionality reduction, maintaining an average normalized root mean square error of 3.5% after the input space was compressed from 169 to 14 easily accessible agronomic variables. Feature importance based on SHapley Additive exPlanations (SHAP) identified land-use dynamics, energy supply chains, animal manure and nutrient management as primary impact vectors, and crop-specific analyses indicated that dataset homogeneity outweighs raw sample volume; both are reported in Section S4 of the Supplementary information. Freshwater Ecotoxicity, Human Toxicity and Photochemical Oxidant Formation remained the least predictable categories, with R 2 below 0.60 in every version of the integrated dataset — a limitation traced to database-specific methodological variability in system boundaries, geographical scope and allocation rules. Deep structural pre-harmonization, functional units beyond pure mass, and parameterized LCA datasets are therefore research priorities. By delivering a parsimonious 14-variable model, the framework bypasses exhaustive supply-chain mapping and supports rapid, sustainability-driven decisions; its outputs are probabilistic estimates, since uncertainty propagation is not yet quantified within the pipeline.