Managing and predicting embodied carbon emissions for ready-mix concrete products using model-agnostic meta-learning technique

Abstract

Ready-mix concrete (RMC) production is a major contributor to upstream carbon emissions in the construction industry. However, the absence of reliable emissions data, coupled with inconsistencies in reporting practices, presents significant challenges for stakeholders in effectively identifying and managing carbon hotspots across regions. Thus, this study employed a web crawling technique to compile a high-quality dataset of 59,412 Environmental Product Declarations (EPD) of RMC products in North America, then utilized the Model-Agnostic Meta-Learning (MAML) algorithm to enhance the embodied carbon emissions prediction for these products. The model was trained using three datasets related to material use, resource consumption, and waste generation as base learners in the United States (U.S.). Then, we tested the model with a new dataset from Canada containing unseen features to evaluate its generalization capability under varying environmental and technological scenarios in RMC production. The results showed that the proposed task-oriented MAML model outperformed the base learners, achieving an R2 score of 0.902 for new task prediction, compared to scores of 0.759, 0.689, and 0.687 for the respective base learners. Furthermore, the MAML model exhibited 25 %–40 % reductions in MAE, RMSE, and MAPE relative to the base learners, highlighting its predictive performance in analyzing multi-task cases. Finally, a web-based platform framework incorporating the trained MAML model is proposed to support stakeholders in managing carbon emissions and to serve as a tool for validating EPD documents for RMC products. The findings of this study provide valuable insights to advance decarbonization efforts within the construction industry. © 2025 The Authors