J Environ Manage. 2026 Jul 5;413:130402. doi: 10.1016/j.jenvman.2026.130402. Online ahead of print.
ABSTRACT
The cement industry contributes approximately 7-8% of global anthropogenic CO2 emissions, yet accurate cradle-to-gate embodied carbon estimation requires plant-level inventory data that are largely unavailable across developing and emerging economies. This data scarcity constrains global benchmarking and the implementation of emerging embodied carbon regulations. This study proposes a STIRPAT-grounded hybrid machine learning-life cycle assessment (ML-LCA) framework for estimating national-scale cement embodied carbon using exclusively publicly available macroeconomic data. GDP per capita, population, and temporal indicators are used to predict the clinker-to-cement ratio (CCR), which is subsequently propagated through a technology-stratified, process-based LCA model enforcing stoichiometric and thermodynamic constraints across A1-A3 stages. Among seven candidate algorithms, Gradient Boosting was selected for its smooth non-linear approximation and LCA integration suitability. SHAP analysis confirms GDP per capita as the dominant CCR driver, with contributions directionally consistent with established technology diffusion theory, ensuring model transparency. Validation across 18 economies through statistical metrics, residual diagnostics, country-level diagnostic benchmarking, Leave-One-Country-Out (LOCO) cross-validation, and three independent literature-benchmarking countries (Pakistan, Mexico, Spain) confirms physically plausible and externally consistent outputs ranging from 0.53 to 0.97 kg CO2/kg cement. A central methodological contribution is the ability to estimate the clinker-substitution decoupling effect at the country scale using only macroeconomic inputs, in contexts where plant-level LCA inventory data are unavailable. Conventional LCA already separates process, energy, and material composition contributions when inventory data are present; the present framework extends this separation to data-scarce national contexts. At the system level, an Environmental Kuznets Curve-type pattern is qualitatively reproduced when model outputs are aggregated across countries, providing a coherence check on the framework as a whole. Out-of-country generalisation is assessed using Leave-One-Country-Out (LOCO) cross-validation as the primary protocol (mean fold RMSE 0.077; 12 of 18 folds below RMSE 0.10), with a forward-chaining temporal split as a complementary diagnostic. The framework is operationalised through an interactive decision-support interface, offering a scalable, transparent baseline for embodied carbon benchmarking, policy screening, and net-zero pathway evaluation in the global cement sector. The framework is positioned as a screening-level reference for data-scarce contexts, complementary to plant-level LCA and Environmental Product Declarations where these are available.
PMID:42402234 | DOI:10.1016/j.jenvman.2026.130402
