Machine Learning-Assisted Life Cycle Assessment and Multi-Objective Optimisation of Low-Carbon Concrete Mixtures

Abstract

Concrete’s embodied carbon is dominated by Portland cement, yet conventional life cycle assessment (LCA) is too slow and data-hungry to guide routine mixture design. This study develops and tests a reproducible computational framework that couples LCA, machine learning (ML), uncertainty quantification, explainable AI and evolutionary optimisation for low-carbon concrete mixtures. No public dataset links mixture proportions, supply-chain context, LCA inventories and strength, so the framework is verified on a literature-parameterised synthetic benchmark of 2,400 mixtures whose strength and global warming potential (GWP) are generated by known equations. Results on the benchmark demonstrate internal consistency, not real-world accuracy; for GWP, R² = 0.998 is structurally expected. Empirical evidence is added for strength only, using the public Yeh dataset (1,030 records, 427 distinct mixtures). On these real data the model ranking changed: tree-based learners generally outperformed quadratic ridge, and the ensemble reached a test RMSE of 4.5 MPa (R² = 0.91). Record-level cross-validation understated the error for unseen mixtures by 17%–22%, and surrogates trained on the synthetic data preserved the rank order of measured strengths (Spearman 0.87) but not their calibration. Within the benchmark’s assumptions, a surrogate-optimised 46.5 MPa mixture had a modelled cradle-to-gate GWP 49.5% below a conventional reference (41.6% below an optimised OPC-only design, where OPC is ordinary Portland cement). The qualitative result held across low-, reference- and high-carbon emission-factor scenarios (46%–53%) and, in direction, under an optimisation constrained by an empirical strength model, but its magnitude depended strongly on the cap on supplementary cementitious materials (SCMs; 33%–58% for caps of 40–70%). Carbon has not been validated against environmental product declaration (EPD) data, and durability enters only through water-to-binder ratio (w/b) and binder limits; the pipeline is intended to be re-calibrated with experimental and EPD data.