Year: 2026
Type: book-chapter
Abstract: The construction sector contributes to 38% of worldwide carbon dioxide (CO2) emissions associated with energy use, with a significant share of a building's embodied emissions stemming from its structural components. Due to the long lifespan of a building's structure, designers must navigate the challenges posed by the long-term implications of their choices. Selecting appropriate structural systems and materials is vital, as it significantly reduces material consumption, waste and the overall environmental impact of a building. Methods for assessing the greenhouse gas (GHG) emissions embodied in structural systems are varied, and these emissions are rarely calculated or evaluated during the early design stage, where system descriptions and material choices are often more conceptual than precise numerical values. Thus, this chapter employs a machine learning (ML) algorithm (random forest classifier) to identify patterns in the GHG emissions of structural systems and to interpret and understand the predictions using the building's features. The considered features are total floor area, use and subuse types of the building, structural and roofing systems, site location and year of completion. As most of the data analysed is descriptive, various encoding algorithms – one-hot encoding, target encoding and label encoding – were tested, and their impact on the prediction model was examined. Mutual information scores were used to quantify the shared information between the considered features of buildings to understand their relationships, interactions and impact on the embodied GHG emissions. Additionally, sequential model-based optimisation (SMBO) was used to identify the optimum set of hyperparameters of the random forest classifier, enhancing the performance of the developed classification models. An examination of the relationships between the building input features reveals the following correlations: the structural system (e.g. struct_system and roof_system) are related to the project's site (country), while the building's total floor area (total_area) strongly correlates with the specified use of the building (subuse_type). These dependencies are suggested by the data and identified by mutual scores. Assessing the feature importance factors from the developed random forest classifier, country emerges as the most significant, indicating the place-dependency of the examined embodied GHG emissions. Structural and roof systems, building subuse type and total floor area are notable contributors to embodied GHG emissions. Furthermore, the analysis of feature importance highlighted specific structural systems, such as massive brick and massive concrete, along with large floor areas, as key contributors to GHG emissions. This underscores the critical influence of material choice and building size on the environmental impact. However, it is important to note that the encoding approach used for the categorical data influenced the features' importance, adding complexity to interpreting the results. For example, label encoding tended to assign greater significance to features with fewer unique labels, such as roof system and year of construction, indicating the effect of ordinality on the classifier learning pattern and prediction outcomes. Building on these findings, integrating ML into understanding and identifying embodied GHG emission patterns will help engineers and designers make informed decisions early in the design phase, contributing to data-driven low-carbon structural design. The developed explainable ML models provide transparency and understandability, enabling users to identify and mitigate biases in ML-based decision-making.
Author: Ghada Karaki
Institution: University of the West of England
Language: English
Subfield: Building and Construction
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7161571809