Machine Learning-Assisted Assessment of Low-Carbon Building Materials for Sustainable Construction Engineering

Abstract

Low-carbon building materials are central to reducing embodied emissions in construction, yet their assessment is complicated by coupled environmental, mechanical, durability, and cost criteria that are measured at different scales and with unequal reliability. Existing life-cycle assessment and material-selection procedures often treat carbon intensity as a scalar attribute and separate it from the statistical uncertainty of property prediction, while machine-learning studies frequently optimize predictive accuracy for a single mechanical response without embedding engineering feasibility or life-cycle boundaries. This manuscript develops an integrated, uncertainty-aware assessment framework that links material and process descriptors to low-carbon performance indicators through surrogate models, calibrates predictive uncertainty, and converts probabilistic outputs into robust design decisions. The framework combines Gaussian-process and ensemble learning for small and heterogeneous construction-material data, multi-objective optimization for carbon, strength, durability, and cost, and constraint-based decision rules that require acceptable probability of compliance rather than point predictions. It also specifies the physical meaning of key variables, including clinker substitution, recycled aggregate porosity, activator demand, thermal conductivity, and fabrication waste, and explains how these variables alter both environmental burdens and engineering reliability. The main contribution is not a claim of empirical superiority but a coherent methodological structure for assessing low-carbon building materials when data are sparse, heterogeneous, and consequential. The framework provides a defensible basis for comparing material alternatives, identifying trade-offs, and exposing the conditions under which machine-learning predictions should not be used.