Interpretable Machine Learning Framework for Embodied Carbon Estimation in Reinforced Concrete Wall Elements.

Abstract

Buildings represent a significant proportion of total anthropogenic greenhouse gas emissions, as operational carbon is gradually reduced by energy efficiency in design. Therefore, embodied carbon produced during the extraction of materials, transportation, manufacture, assembly, disassembly, and disposal of structural components is receiving increasing attention in literature. Conventional Life-Cycle Assessment (LCA) techniques, while comprehensive, have proven to be slow and tedious and cannot be applied repeatedly during early-stage design. This thesis addresses the above gap through the development of a machine learning model capable of predicting the embodied carbon intensity of Reinforced Concrete (RC) walls based on early-stage design inputs.

Synthetic data of a considerable size was created by systematically permuting ten design variables related to wall configuration (geometrical properties, type and strength of concrete used, ratio of reinforcement), and transport distances, along with the results of life cycle carbon calculation using the Inventory of Carbon and Energy (ICE) emission factors and cradle-to-grave system boundary defined by EN 15804.

The dataset includes a variety of structural configurations that can be considered realistic for low-to mid-rise structures in Finland. An XGBoost regression model was trained on the dataset and assessed using regression metrics and five-fold cross-validation. The model interpretability was analyzed using various methods of explanation, namely, Global and Local SHAP, LIME-based local interpretation, Partial Dependence Plots (PDP), Individual Conditional Expectation curves (ICE), and Two-Way PDPs. The above methods were selected to achieve sufficient model interpretability and avoid black-box estimation.

As a result, an XGBoost model was obtained with nearly perfect predictive performance demonstrated across all cross-validation folds. Model interpretability revealed wall thickness to be the most influential design variable, followed by the compressive strength of concrete and reinforcement ratio. This was expected since the above variables define material volumes, as well as associated carbon emission rates and their domination was observed across all methods of model explainability. Transport-related variables exhibited systematic impact to the extent lower than material-related, while wall length and wall height proved to be relatively unimportant for predicting embodied carbon in units of area.

This study proves the applicability of an interpretable ML model for rapid evaluation of multiple alternative configurations at an early stage of design, without the need for complete LCA calculation for each. The current study contributes to the body of literature, as it focuses on predicting embodied carbon on a structural element level rather than material-level and whole-building prediction. For future research, validation of the developed model against empirically collected data and expansion of the framework for predicting the embodied carbon of other RC elements are recommended.