Explainable AI-Driven Analysis of Construction and Demolition Waste Credit Selection in LEED Projects

Abstract

Selecting Construction and Demolition Waste (CDW) credits in LEED-certified projects is essential for sustainable building management, often requiring specialised expertise and contextual sensitivity. However, existing studies provide limited analytical insight into why certain CDW credits succeed or fail across different project contexts, and no explainable AI–based framework has been proposed to support transparent credit decisioning. This gap underscores the need for a data-driven, interpretable approach to CDW credit evaluation. This study proposes an explainable artificial intelligence (XAI)-based model to support CDW credit selection and to identify the key factors influencing credit performance. A dataset of 407 LEED green building projects was analysed using twelve machine learning (ML) algorithms, with the top models identified through Bayesian optimisation. To handle class imbalance, the SMOTE was utilised. Results showed that MRc2 and MRc4 credits had high predictive performance, while MRc1.1 and MRc6 credits exhibited relatively lower success rates. Due to data limitations, MRc1.2 and MRc3 were excluded from analysis. The CatBoost model achieved the highest performance across MRc1.1, MRc2, MRc4, and MRc6, with F1 scores of 0.615, 0.944, 0.878, and 0.667, respectively. SHapley Additive exPlanations (SHAP) analysis indicated that the Material Resources feature was the most influential predictor for all credits, contributing 20.6% to MRc1.1, 53.4% to MRc2, 36.5% to MRc4, and 22.6% to MRc6. In contrast, the impact of design firms on credit scores was negligible, suggesting that although CDW credits are determined in the design phase, these firms did not significantly influence the decision process. Higher certification levels improved the performance of MRc1.1 and MRc6, while their effect on MRc2 and MRc4 was limited. This study presents a transparent and interpretable XAI-based decision-support framework that reveals the key sustainability drivers of CDW credit performance and provides actionable guidance for LEED consultants, designers, and decision-makers.

1. Introduction

Green building (GB) is a comprehensive approach that integrates sustainability throughout the design and operational phases, aiming to minimise environmental impacts [1,2,3]. GBs enhance ecological and human health by promoting energy efficiency, reducing carbon emissions, and improving indoor air quality [4,5]. Their primary goals include achieving high environmental performance, ensuring user well-being, and generating long-term economic benefits [6]. An essential component of this process is the effective management of Construction and Demolition Waste(CDW), which is critical to advancing sustainability objectives. Sönmez and Maçka Kalfa [7] highlighted significant gaps in definitions, scope, standards, and data quality regarding CDW legislation and policies under the European Union’s Waste Framework Directive (WFD). Globally, the construction sector consumes approximately 40% of raw materials, generates around 40% of solid waste, and accounts for 30% of energy-related greenhouse gas emissions [8]. These figures underscore the critical importance of CDW management in achieving sustainable development goals.

Green Building Rating Systems (GBRS) assess environmental performance based on criteria such as energy efficiency, material selection, waste management, and water consumption [9,10]. They certify projects through performance-based evaluation, assuring investors [11]. In the design phase, architects must align strategies with certification requirements, including energy, water, and indoor environmental quality [12]. LEED, the most widely adopted system, is internationally recognised [13]. However, the absence of a universal standard and regional variations in criteria present a research gap, particularly in CDW management [5,14].

Experts often depend on personal knowledge and experience when making decisions, such as selecting credits in GB projects [12]. However, the lack of a structured guide makes credit selection time-consuming and labour-intensive for architects, reducing the feasibility of GB projects. Therefore, some studies suggest using traditional statistical methods to analyse variations in credit performance under different conditions, with a focus on credit selection. For instance, Yu et al. [15], Chi et al. [16], and Daoud et al. [17] conducted statistical analyses of CDW credit performance based on previous projects. Comparative GBRS studies have examined system goals, scoring structures, and implementation costs [18], but have not integrated AI-based decision-support tools or CDW credit performance analysis. This shows that XAI-supported approaches remain unexplored in the context of CDW management. Other works have focused on specific aspects, such as the economic dimension of transportation credits [19] or the economic impacts of LEED credits using GIS and regression models [20]. However, they excluded AI-driven decision support and CDW credits. While valuable, such studies remain limited for developing credit selection strategies across diverse contexts, and existing decision support systems still lack explainability [21].

The complex relationships between increasing GBRS criteria and multiple impact factors have heightened the need for artificial intelligence (AI) techniques, especially machine learning (ML). ML, a subfield of AI, is an approach that develops models capable of analysing complex structures in high-dimensional data, extracting task-specific features, and automating decision-making and prediction processes by learning from past data [22,23,24]. As a result, it can be effectively utilised across a broad range of domains, from text classification to planning [25].

In recent years, ML-based AI applications have significantly improved CDW management within GBRS. For example, embodied carbon, material selection, and waste plans can be digitised through BIM integration and databases [21,26]. XGBoost with SHAP has supported credit selection in design phases [12], while SVR linked LEED building distribution with socio-institutional factors [27]. Genetic algorithms and BIM reduced waste at the building level [28], and evolutionary optimisation in Norway decreased embedded carbon waste by 14.5% [29]. These cases demonstrate how AI models support CDW management, both directly and indirectly, at multiple scales.

However, existing AI models cannot be generalised to cover all GBRS credit categories, and their impact on decision support processes is limited due to factors such as data imbalance [12,27,30,31], lack of LCI [29], and contextual constraints [21,28,32,33]. Although Xu et al. [34] and Mansouri et al. [35] demonstrated the predictive potential of ML for LEED performance, and Payyanapotta & Thomas [36] showed applications in prefabrication, none addressed CDW credits or explainable AI-based decision support. Additionally, the focus on CDW is limited to specific material categories, indicating a partial perspective that does not align with the circular economy approach [21]. The literature emphasises that the impact of GB projects on CDW management remains underexplored [37,38,39,40,41]. In this context, CDW-related credits represent a crucial area for future research, and improving them through data-driven decision-making systems can directly support sustainable building policies.

The growing advancement and increasing complexity of AI systems have raised significant concerns regarding the explainability and accountability of model-generated decisions [42,43]. These concerns arise from the inherently ‘black-box’ behaviour of many AI models, which makes it difficult for users to understand how predictions are made [44]. As AI applications continue to expand within the CDW domain, ensuring transparency in decision-making processes has become increasingly important [45]. This has driven the development of explainable artificial intelligence (XAI) approaches that make decision logic more transparent and interpretable [44,46]. XAI enhances the interpretability of model outputs, allowing decision-makers to understand, evaluate, and trust AI-generated recommendations [47,48,49].

Within this context, XAI offers a quantitative insight into how sustainability indicators influence CDW credit performance through SHAP analysis, the method exclusively used in this study, therefore enhancing data-driven assessment processes within the GBRS framework [12,50,51]. As shown in previous research [50,52], SHAP-based explanations identify which indicators increase or decrease credit success, assisting architects, LEED consultants, and decision-makers in conducting more consistent and rational evaluations. To our knowledge, this is the first study to apply a SHAP-based XAI technique to assess CDW credit performance within the GBRS/LEED framework. In line with recent advances in AI for sustainability assessment, this study explicitly incorporates machine learning and explainable AI (XAI) into its research, aiming to offer transparent, data-driven support for CDW credit decisions.

This study utilises XAI models to examine the factors influencing the selection of CDW credits in LEED-New Construction (NC) v3-certified projects across Turkey, Germany, and Spain. An ML-based decision-support model was developed using a dataset of 407 projects, with the model’s interpretability ensured through SHAP analysis. The study sets out three main objectives: (1) to predict CDW credit selections with ML models, (2) to identify the key factors influencing credit selections and their significance using SHAP analysis, and (3) to improve transparency in the decision-making process by evaluating the model’s accuracy and reliability using ML and XAI techniques. To ensure an AI-oriented framing, the research questions explicitly focus on how ML and XAI can enhance the accuracy, interpretability, and transparency of CDW credit decision-making. Previous studies have demonstrated that CDW performance in LEED projects is affected by a range of contextual and design-related factors, including certification level, building type, architectural or design firm, economic conditions, environmental sustainability indicators, and climate. However, these factors have not been comprehensively examined within a single analytical framework in existing research. Consequently, research questions were developed to explore the combined influence of these factors on CDW credit decisions, based on the key determinants listed in

Table 1. CDW–related credits in the LEED-NC v3.

Overall Assessment Framework		Credits Related to CDW
Performance Category	Attainable Points			Attainable Points
Sustainable sites (SS)	26	MRp1	Storage and collection of recyclables	prerequisite
Water efficiency (WE)	10	MRc1.1	Building reuse—maintain existing walls, floors, and roof	3
Energy & atmosphere (EA)	35	MRc1.2	Building reuse—maintain existing interior non-structural elements	1
Materials & resources (MR)	14	MRc2	Construction waste management	2
Indoor environmental quality (IEQ)	15	MRc3	Materials reuse	2
Innovation (IN)	6	MRc4	Recycled content	2
Regional priority (PRC)	4	MRc6	Rapidly renewable materials	1
Total	110			11

. The primary research questions of the study are as follows:

How do a wide range of factors—including certification level, building type, architectural firm, climate, sustainability, and economic conditions—impact CDW credit decisions in LEED projects?

How can explainable artificial intelligence (XAI) models provide optimised decision support to architects and decision-makers in this selection process?

These questions centre on AI, especially XAI, as the study’s main analytical perspective. In line with these research questions, machine learning and XAI models analysed the key factors influencing CDW credit selections and their impacts. This study fills a significant gap in the literature by providing a transparent, data-driven decision-support framework for CDW credit selection using explainable AI, thereby increasing the accuracy and interpretability of sustainable building assessments.

2. Research Methodology

2.1. Research Framework

Due to the limited number of LEED-NC v4 certified projects, this study focused on LEED-NC v2009 [53]. According to USGBC [54], Turkey, Germany, and Spain rank among the top 10 EU countries in LEED-certified gross floor area [55], with 1129, 1591, and 1595 projects, respectively, as of March 2024. This study analysed 407 LEED-NC v3 projects from the USGBC database, focusing on new construction projects in which CDW credit decisions are primarily made during the design phase.

The research framework of this study, shown in Figure 1, includes three main stages: (1) data preparation, (2) model development, and (3) model evaluation using SHAP-based interpretability analysis. The framework consists of three main stages. In the first stage (data preparation), relevant features were identified, data were collected, and a structured database was created through preprocessing, feature engineering, and SMOTE-based data balancing. In the second stage (model development), multiple ML algorithms were applied. Hyperparameter optimisation was conducted using the Bayesian Search algorithm, with eight feature categories serving as model inputs and each CDW credit represented as a binary target. The effects of SMOTE, hyperparameter optimisation, and threshold adjustment were evaluated by comparing results before and after optimisation. In the third stage (model evaluation), separate models were developed for each CDW credit. The top-performing model was selected based on the highest F1 score, and the Best Threshold method was applied to balance sensitivity and specificity. Finally, SHAP analysis identified the most influential features and their contributions to credit selection. The ultimate goal of the framework is to provide transparent, reliable, and explainable decision-support tools for designers, architects, and decision-makers.

2.2. Data Preparation

The data preparation process involves two main steps, preprocessing and feature engineering, that improve data quality by reducing noise and enabling more accurate ML predictions. LEED criteria consist of required prerequisites and optional credits. This study focuses on the optional CDW credits listed in Table 1 and examines one CDW-related prerequisite, along with six optional CDW-related credits, within the LEED framework.

The dataset comprises 407 LEED-NC v3-certified projects from the USGBC database, with 29 features: 22 independent features, 6 target features (related to CDW credits), and 1 identifier. Each observation represents a project. Two approaches are common in the literature for evaluating credit performance: (1) ratio of earned to possible points [56,57,58,59], and (2) binary classification above or below the mean score [12,30,60,61]. Success rates were quantified using the Credit Achievement Degree (CAD) ratio, where CO (credit obtained) represents the points earned by the project and TC (total credit or total possible points) denotes the maximum points available for that credit. The second approach classifies credit performance as a binary variable, with projects scored above or below the average.

$${Score}_{ratio}=CAD={\displaystyle \frac{CO\left(pointsoflayerobtained\right)}{TC\left(totalpointsofthelayer\right)}}$$

(1)

$${Score}_{binary}=\left\{\begin{array}{cc}1,& ifScore\ge \mu \\ 0,& ifScore<\mu \end{array}\right.$$

(2)

where μ represents the mean score of the credit categories.

In this study, the ratio-based approach was applied for continuous input features, whereas the binary approach was used for CDW credit targets, consistent with prior research.

Based on prior studies, input factors were grouped into 8 categories (

Table 2. Definitions of key influencing features.

Category/Factor	Feature	Feature Type	Feature Encoding	Sources
Certification level	Certified	Categorical	Label encoding (0/1/2/3)	[53]
	Silver
	Gold
	Platinum
Building type	Residential buildings	Categorical	One-hot encoding [1, 0, 0], [0, 1, 0], [0, 0, 1]	Adapted from [12,53,59]
	Science, education, culture, and health buildings (SECH)
	Commercial, industrial, and office buildings (CIO)
Architectural firm	Whether the architectural design firm ranks among the top 100 (WE100)	Categorical	Binary encoding (0/1)	[65]
Economic factor	GDP per capita	Numerical	Continuous feature	[66]
Sustainable factor	GER (Green Efficiency Ratio)	Numerical	Continuous feature	Derived based on [53]
Climate factor	Average temperature	Numerical	Continuous feature	[67]
	Summer temperature
	Winter temperature
	Temperature difference
	Annual precipitation
	Annual sunshine time
	Average wind speed
LEED credit categories	Sustainable sites	Numerical	Continuous feature	[53]
	Water efficiency
	Energy atmosphere
	Material resources
	Indoor environmental quality
	Innovation
	Regional priority credits
Project area	Gross floor area	Numerical	Continuous feature	[53]

), resulting in 22 independent features. The factors used in this study constitute the inputs for the machine learning and XAI models. Based on previous research, the model included 22 factors across eight categories that could influence CDW credit decisions in LEED-NC v3 projects. Firstly, certification level is a key determinant of CDW performance; indeed, Chi et al. [16] found no significant difference at the platinum level, but differences become more evident at lower levels. Furthermore, building type is another critical factor influencing sustainability priorities and credit selection behaviours [12,59]. The different design and material requirements of residential, SECH, and commercial/office buildings explain differences in CDW performance. Furthermore, design firms play a crucial role in the success of sustainable building projects [62,63,64]. Therefore, whether an architectural design office ranks among the top 100 firms (WE100) was added to the model as an indicator of the firm’s capacity and expertise. Additionally, high costs [39,40] and technological and economic constraints [16] complicate CDW management. Because the maturity of the green building market in more economically developed regions can influence CDW loan performance [7,38], GDP per capita was included as a proxy for economic conditions.

GER is an indicator of environmental sustainability performance, and a high GER value signifies urban structures that prioritise environmental performance, which can be linked to the success of sustainability-focused loans like CDW. Additionally, climate significantly influences building design and sustainability practices; therefore, CDW requirements can vary across GBRSs depending on local climate and design choices [10,68]. Consequently, seven climate indicators commonly used in the literature were included in the model. Furthermore, credit scores across other LEED categories provide comprehensive information about a project’s sustainability profile; Ma & Cheng [31] even demonstrate that performance in one category can be linked to performance in others through integrated design approaches and project features. Lastly, since the total construction area indicates the scale of the project [55], the complexity of the construction process and the increased CDW in larger projects imply that CDW strategies may differ with total area.

Climate-related features were obtained from the Meteostat Python library [65], representing monthly averages for 2010–2024. For cities with missing values, data from the nearest meteorological station were used. Economic indicators were represented by gross domestic product (GDP) per capita from the Eurostat [66] database, with averages from 2013 to 2024 reflecting long-term national trends. To capture the influence of architectural design firms, projects were matched to the World Architecture 100 (WA100) ranking [67]. A binary feature was created, assigning a value of 1 to projects designed by WA100 firms and 0 to all other projects.

2.2.1. Data Preprocessing

In the preprocessing stage, several steps were applied to prepare the dataset for machine learning. Outliers and inconsistent observations were identified using histogram plots, and a logarithmic transformation was applied to approximate normality. Categorical features were encoded as follows: binary for binary features, label for ordinal features, and one-hot for non-ordinal categories. Numerical features were normalised with StandardScaler. Since a class imbalance was observed between projects that achieved and those that did not achieve CDW credits, the Synthetic Minority Over-sampling Technique (SMOTE) was applied to the training set (Figure 2).

In this study, a 60% imbalance threshold was considered critical, following the work of Chawla et al. [69] and He and Garcia [70]. For the target features MRc1.1, MRc1.2, MRc3, and MRc6, synthetic samples were generated for the minority group of projects that earned CDW points. For MRc2 and MRc4, synthetic data were generated for the minority group of projects that did not earn CDW points. This approach ensured a balanced training set.

2.2.2. Feature Engineering

During feature engineering, existing features were analysed, and a new derived indicator, the Green Efficiency Ratio (GER), was created. GER was calculated as the ratio of a project’s total LEED score to its gross floor area, representing sustainability efficiency. Distribution analyses revealed that GrossFloorArea, Precip_mm, Sunshine_hours, and GER exhibited significant deviations from normality and numerous outliers; therefore, a logarithmic transformation was applied, which improved their distributions. Hyperparameter optimisation was conducted with five-fold cross-validation and the Bayesian search algorithm. Feature iteration was used to identify features that enhanced predictive performance, and SHAP-based feature selection was applied to improve interpretability.

3. Model Selection and Development Process

This section applied and optimised several machine learning algorithms to analyse factors influencing CDW credit selections in LEED-NC v3 projects using XAI models. The aim was to provide designers and decision-makers with transparent, data-driven decision support. The dataset was split into training (80%) and testing (20%) sets, followed by hyperparameter tuning and the best-threshold method. To ensure complete experimental consistency, the same 80–20 train–test split (random_state = 42) was applied across all algorithms, yielding identical training and test sets for each model. Test data were kept independent and used only for the final performance evaluation. The input data consisted of 22 independent features grouped into eight categories, while the target comprised six binary features indicating whether each CDW credit was achieved (e.g., [1, 0, 0, 1, …]). Since the target variables (CDW credits) are binary, the Pearson correlation coefficient was preferred. Pearson is equivalent to the point-biserial correlation for binary variables [71] and provides the most appropriate representation of the relationships among this data type. Figure 3 presents the correlation matrix of target features, with values ranging from −1 (strong negative) to +1 (strong positive).

The correlation among target features ranges from 0.25 to −0.24, indicating weak relationships. Therefore, these correlations were considered negligible in the analysis, and six separate single models were developed instead of a multi-target model.

3.1. Algorithm Selection and Model Configuration

This study employed 12 supervised ML algorithms to predict CDW credit acquisition in LEED-certified projects. These include linear models (Ridge and Logistic Regression), tree-based methods (Decision Tree, Random Forest, Extra Trees), instance-based learning (K-Nearest Neighbours), Gaussian Naive Bayes and Support Vector Machines (SVMs), and gradient-boosted decision trees (GBM, XGBoost, LightGBM, CatBoost). The inclusion of these algorithms enabled performance comparisons across distinct learning paradigms linear, nonlinear, ensemble-based, distance-based, probabilistic, and boosted models providing a comprehensive assessment of a heterogeneous and imbalanced tabular dataset.

To address potential multicollinearity, ridge regression was employed, which regularises model complexity through the L2 norm [72]. Logistic regression is a widely used classification method for binary outcomes [73], modelling event probability and estimating the effect of each predictor through odds ratios [74]. Naive Bayes is a simple probabilistic classifier based on Bayes’ Theorem that assumes feature independence and selects the class with the highest posterior probability [75]. k-NN is a non-parametric model that classifies new observations based on their proximity to the most similar labelled samples, typically using Euclidean distance [76,77]. DT regression partitions the feature space into homogeneous regions [78]. Extra Trees extends this by introducing higher randomisation in split selection [79], while Random Forest aggregates multiple trees through bootstrap sampling, reducing variance and overfitting [78,80]. SVM identifies the hyperplane that maximises the margin between classes, relying on support vectors to determine the decision boundary [81].

Gradient boosting models, including AdaBoost, GBM, XGBoost, LightGBM and CatBoost, iteratively combine weak learners to minimise residual errors and consistently deliver strong predictive performance on heterogeneous, high-dimensional tabular datasets [82,83]. AdaBoost reweights misclassified samples to improve the performance of weak learners [84]. GBM fits trees sequentially by following the gradient of the loss function [85]. XGBoost incorporates regularisation and efficient handling of missing data to reduce overfitting [86]. LightGBM accelerates training through histogram-based splitting and leaf-wise tree growth [87]. CatBoost uses ordered boosting and native handling of categorical variables to prevent prediction shift and improve accuracy [88].

The modelling process was implemented in Python 3.10 (PyCharm 2022.2.3) using open-source libraries such as scikit-learn, matplotlib, and SHAP. All computations were performed on a Lenovo laptop equipped with an Intel Core i7-9750H processor (2.60 GHz) and 16 GB of RAM.

3.2. Hyperparameter Tuning

To identify the key factors affecting CDW credit selections, ML models were developed for each credit, and hyperparameter optimisation was applied to improve their performance. Bayesian optimisation was employed to identify the parameter combinations that maximise model success. Using the Skopt library, this process was guided by the Expected Improvement (EI) acquisition function, which balances exploration and exploitation by prioritising hyperparameter combinations with the highest probability of outperforming the current best result [89]. The mathematical definition of this function is presented in Equation (3).

$$EI\left(x\right)=\left\{\begin{array}{cc}\left(E\left[f_{min}\right]-f\left(x\right)\right)P\left(f\left(x\right)\le f_{min}\right)& if\sigma \left(x\right)>0\\ \hfill 0\hfill & otherwise\hfill \end{array}\right.$$

(3)

Here, f(x) represents the model’s prediction at the candidate point, $f_{min}$ is the best-observed value so far, and σ(x) denotes the uncertainty at that point. The probability term is computed using the cumulative normal distribution function. The best hyperparameter combinations for each model were identified using Bayesian optimisation with the EI acquisition function, after which the models were retrained and compared based on their F1 scores.

Table 3. Optimal hyperparameter values for the machine learning models.

Model Abbreviations	Model	Parameters
Ridge	Ridge Classifier	Alpha 0.01–10.0 (log-uniform)
LogReg	Logistics Regression	C 0.01–10.0 (log-uniform)
DT	Decision Tree	max_depth 3–10
ET	Extra Trees	n_estimators 50–200, max_depth 3–10
k-NN	K-Nearest Neighbours	n_neighbors 3–15
NB	Gaussian Naive Bayes	var_smoothing 1 × 10−9 to 1 × 10−5 (log-uniform)
RF	Random Forest	n_estimators 50–200, max_depth 3–10
SVM	Support Vector Machine	C 0.01–10.0 (log-uniform), gamma 0.001–1.0 (log-uniform)
GBM	Gradient Boosting Machine	n_estimators 50–200, learning_rate 0.01–0.5 (log-uniform), max_depth 3–7
XGBoost	Extreme Gradient Boosting	n_estimators 50–200, learning_rate 0.01–0.5 (log-uniform), max_depth 3–7
LightGBM	Light Gradient Boosting Machine	n_estimators 50–200, learning_rate 0.01–0.5 (log-uniform), max_depth 3–7
CatBoost	Categorical Boosting	iterations 50–500, learning rate 0.01–1.0 (log-uniform), depth 3–7, l2 leaf regularisation 1–5 (log-uniform), bagging temperature 0.5–1.5

presents the optimal hyperparameter values obtained for all machine learning algorithms.

Hyperparameter optimisation methods are especially important for algorithms that require extensive tuning. In this study, Bayesian hyperparameter optimisation (BayesSearchCV) was used to find the best hyperparameters for each model. The hyperparameter ranges listed in Table 3 were deliberately set broadly to allow the models to explore a vast solution space; for scale-sensitive parameters (such as C, α, learning_rate, and var_smoothing), a log-uniform distribution was employed. All models were integrated into a processing pipeline comprising SMOTE and standardisation steps (SMOTE → StandardScaler → classifier), and Bayesian optimisation was performed with 50 search iterations and 5-fold cross-validation across all algorithms. The 50 iterations refer to the number of hyperparameter combinations evaluated by BayesSearchCV and do not correspond to model hyperparameters such as iterations or n_estimators. Due to class imbalance, the optimisation objective was set to maximise the F1 score, which jointly captures precision–recall performance (scoring = ‘f1’). For CatBoost, the iterations parameter was tested over 50–500, while tree-based models evaluated n_estimators over 50–200. No additional analysis of hyperparameter importance was performed, and the optimisation process depended solely on selecting the values that best satisfied the objective function.

Among the boosting models listed in Table 3 are AdaBoost, GBM, LightGBM, XGBoost, and CatBoost. These algorithms are configured through several key hyperparameters that directly affect model learning behaviour and generalisation performance. The learning_rate controls each weak learner’s contribution and prevents overly aggressive updates that could lead to overfitting. The n_estimators parameter (or iterations in CatBoost) determines the number of boosting steps and thus the model’s overall learning capacity. The max_depth (or depth in CatBoost) parameter limits tree complexity by restricting the maximum split depth, helping to balance overfitting and underfitting. CatBoost also uses l2_leaf_reg, an L2 regularisation term on leaf values to reduce model variance, and bagging_temperature, which influences sampling randomness and adjusts the bias–variance trade-off. These descriptions clarify how the boosting models in Table 2 function and ensure the reproducibility of the experimental setup.

3.3. Model Evaluation Metrics

The model’s performance was evaluated using six metrics. Each prediction was classified as True Positive (TP), True Negative (TN), False Positive (FP), or False Negative (FN). Correctly predicting projects with CDW credits was labelled TP; correctly predicting those without credits as TN; misclassifying projects without credits as credited as FP; and failing to identify credited projects as FN. Threshold-based performance was assessed using the Receiver Operating Characteristic (ROC) curve and its Area Under the Curve (AUC), which summarises the classifier’s discriminatory ability across decision thresholds. Overall performance was measured using accuracy, precision, recall, F1 score, AUC, and Brier score. Accuracy is the proportion of correctly predicted samples; precision is the share of truly credited projects among predicted positives; recall is the share of correctly identified credited projects; and the F1 score is the harmonic mean of precision and recall. The formulas and definitions of these metrics are provided in

Table 4. Model performance metrics and their formulas.

Metric	Formula	Description
Accuracy	${\displaystyle \frac{TP+TN}{TP+TN+FP+FN}}$	The model’s overall predictive performance
Precision	${\displaystyle \frac{TP}{TP+FP}}$	The degree of precision in the model’s correct predictions
Recall	${\displaystyle \frac{TP}{TP+FN}}$	The proportion of actual positive instances correctly identified
F1 Score	${\displaystyle \frac{2TP}{2TP+FP+FN}}$	A useful hybrid metric for imbalanced classes
AUC	${\displaystyle \frac{1}{2}}{\displaystyle \sum _{i=1}^{n-1}}({TPR}_{i+1} + {TPR}_i)({FPR}_{i+1} - {FPR}_i)$	Measures the model’s ability to distinguish positive and negative classes
Brier Score	BS = ${\displaystyle \frac{1}{N}}{{\sum }_{i=1}^N(p_i-y_i)}^2$	An error metric measuring prediction accuracy and reliability

.

The Brier score (BS) measures the reliability and accuracy of a classification model. Here, N represents the number of observations, p represents the model’s predicted probability for the i-th observation, and y represents the actual label (0 or 1). A lower Brier score indicates more reliable predictions. In classification problems, evaluating the F1 score and reliability together is critical.

3.4. SHAP Analysis for Explainability

Explainable Artificial Intelligence (XAI) enhances model transparency and interpretability, strengthening reliability [46,90]. In this study, XAI is employed to clarify credit decisions for CDW, thereby making them more transparent and trustworthy. Among XAI methods, Shapley Additive Explanations (SHAP) provide a detailed analysis of the decision-making processes of machine learning models [90]. Based on game theory [91], SHAP assigns model output to input features based on their contributions. The explanation model is expressed as a linear combination of input features, with coefficients represented by Shapley values (φ_i) [46,92]. SHAP plots visualise the impact of each feature on predictions [46], where the sign of φ_i indicates whether the feature increases or decreases the output, and the magnitude reflects the strength of this effect [93]. Models with higher accuracy provide more reliable estimates of feature importance [94]. Equations (4) and (5) represent this combination and the computation of Shapley values.

$$g(x^{\prime })=f(h(x^{\prime }))={\mathrm{\varphi }}_0+\sum _{i=1}^M{\mathrm{\varphi }}_{İ}{X}_{İ}^{\prime }$$

(4)

$${\mathrm{\varphi }}_i={\sum }_{S\subseteq F\backslash \left\{i\right\}}{\displaystyle \frac{\left|S\right|!\left(\left|F\right|-\left|S\right|-1\right)!}{\left|F\right|!}}\left[f_{su\left\{i\right\}}\left(x_{su\left\{i\right\}}\right)-f_s\left(x_s\right)\right]$$

(5)

Equation (4) describes the model output as the sum of each feature’s contributions. Here, ${\mathrm{\varphi }}_0$ indicates the model’s average prediction value (bias), while ${\mathrm{\varphi }}_i$ represents the contribution of each feature to the model. Equation (5) explains the method for calculating SHAP values. In this equation, S represents a subset of all features, and $f_s\left(x_s\right)$ denotes the model prediction based only on the subset S. The term $f_{su\left\{i\right\}}\left(x_{su\left\{i\right\}}\right)$ indicates the model prediction when the feature i is added. This structure assesses how each feature influences the model output by analysing the presence or absence of other features.

4. Results

4.1. Evaluation Results of ML Models

In the initial phase of the analysis, 12 machine learning models were trained with tuned hyperparameters, and their performance on the imbalanced dataset was evaluated. Model performance was evaluated using classification metrics, including Accuracy, Precision, Recall, F1 score, and Brier score, which varied across the CDW credits. To address the class imbalance in the dataset and improve model performance, SMOTE was applied to all six CDW credits. However, due to insufficient data for MRc1.2 (5 samples) and MRc3 (13 samples), the F1 scores remained unsatisfactory despite SMOTE, hyperparameter tuning, and best threshold methods. Consequently, these credits were excluded from the analysis. The model performance metrics for the remaining four CDW credits are presented in

Table 5. Performance comparison of machine learning models.

		MRc1.1		MRc2		MRc4		MRc6
Model	Metric	Without Smote	With Smote	Without Smote	With Smote	Without Smote	With Smote	Without Smote	With Smote
Ridge	F1 score	0.000	0.522	0.857	0.897	0.796	0.755	0.000	0.250
	Brier score	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN
LogReg	F1 score	0.609	0.609	0.898	0.904	0.846	0.853	0.333	0.333
	Brier score	0.088	0.114	0.137	0.133	0.165	0.179	0.057	0.154
DT	F1 score	0.375	0.222	0.891	0.917	0.847	0.862	0.133	0.615
	Brier score	0.103	0.181	0.106	0.127	0.169	0.164	0.159	0.145
ET	F1 score	0.500	0.560	0.914	0.895	0.859	0.850	0.500	0.545
	Brier score	0.082	0.143	0.139	0.164	0.171	0.171	0.045	0.139
k-NN	F1 score	0.308	0.400	0.815	0.836	0.831	0.829	0.333	0.250
	Brier score	0.103	0.165	0.201	0.213	0.199	0.232	0.066	0.139
NB	F1 score	0.462	0.533	0.832	0.839	0.829	0.841	0.375	0.400
	Brier score	0.115	0.292	0.195	0.199	0.290	0.298	0.477	0.476
RF	F1 score	0.500	0.545	0.935	0.944	0.862	0.871	0.364	0.526
	Brier score	0.084	0.116	0.110	0.115	0.156	0.146	0.050	0.088
SVM	F1 score	0.444	0.545	0.903	0.895	0.857	0.846	0.421	0.571
	Brier score	0.091	0.115	0.129	0.119	0.163	0.174	0.045	0.087
GB	F1 score	0.452	0.571	0.929	0.885	0.871	0.869	0.444	0.615
	Brier score	0.110	0.152	0.095	0.123	0.150	0.142	0.068	0.050
XGBoost	F1 score	0.438	0.533	0.925	0.927	0.867	0.866	0.381	0.533
	Brier score	0.096	0.107	0.105	0.106	0.147	0.145	0.054	0.075
LightGBM	F1 score	0.571	0.522	0.917	0.923	0.857	0.867	0.571	0.500
	Brier score	0.090	0.117	0.106	0.101	0.149	0.140	0.037	0.105
CatBoost	F1 score	0.526	0.615	0.926	0.944	0.871	0.878	0.571	0.667
	Brier score	0.093	0.125	0.099	0.088	0.168	0.136	0.035	0.050

Note: RidgeClassifier does not provide probability outputs; therefore, Brier score optimisation was not computed. Bold values highlight the highest F1 scores and lowest Brier scores.

. A detailed review evaluating all metrics in the table is provided in the Appendix A.

According to Table 5, applying the SMOTE significantly enhanced classification performance, especially in credit categories with notable class imbalance. For instance, in the MRc6 category, the DT model’s F1 score increased from 0.133 to 0.615, a 362.4% increase. Similarly, the Ridge model in the MRc1.1 category improved from 0.000 to 0.522, indicating that a model initially unable to classify achieved meaningful predictive capability. In MRc2, the same model’s F1 score increased from 0.857 to 0.897, representing an absolute improvement of 0.04. A more modest improvement was observed in the MRc4 category with the DT model, where the F1 score increased from 0.847 to 0.862. However, it was also observed that applying SMOTE does not always produce positive results. For example, in the MRc1.1 category, the DT model declined by 0.153 points, while some models showed no change in performance (e.g., the LogReg model in MRc1.1). F1 scores improved in 68.75% of the models, declined in 27.08%, and remained unchanged in 4.17%, suggesting that SMOTE generally contributed positively. These findings indicate that SMOTE’s effectiveness depends on both the credit type and the model’s structural characteristics, with significant improvements observed particularly in imbalanced datasets and in cases with low initial performance.

Model performance was evaluated using two key metrics the F1 score, which indicates classification success, and the Brier score, which measures the reliability and calibration of probabilistic predictions. High F1 scores and low Brier scores signify optimal performance. After applying SMOTE, the CatBoost model consistently achieved the highest performance across all four MR credit categories. Specifically, the F1 and Brier scores for CatBoost on the SMOTE-balanced dataset were 0.615 and 0.125 for MRc1.1, 0.944 and 0.088 for MRc2, 0.878 and 0.136 for MRc4, and 0.667 and 0.050 for MRc6. Notably, in the MRc6 category, the highest F1 score increased from 0.364 (Random Forest, without SMOTE) to 0.667 (CatBoost, with SMOTE), a relative rise of 83.2%. Similarly, in MRc2, CatBoost’s F1 score increased from 0.926 to 0.944, demonstrating that even high-performing models can benefit from improved class balance. These findings highlight CatBoost’s robust classification accuracy and predictive reliability, though optimal model selection may still vary by credit category and dataset characteristics.

4.2. SHAP Analysis Result

Selecting appropriate CDW credits for GB projects is a critical yet challenging process influenced by various factors. Explaining model outputs with SHAP technology helps identify the key factors affecting CDW credit selection and clarifies their influence on the decision-making process. This study conducted an explainability analysis using the TreeSHAP method for the models that achieved the highest performance in each credit category of the CDW. While these models achieved high predictive accuracy, SHAP-based analyses were used to improve the transparency and interpretability of the decision-making process. The results enabled the visualisation of key features influencing model predictions and clarified their impact on credit scores. For example, the CatBoost model demonstrated the best performance for MRc1.1 (building reuse—maintain existing walls, floors, and roof), MRc2 (construction waste management), MRc4 (recycled content), MRc6 (rapidly renewable materials) credits, and SHAP analysis identified the features with the most significant impact on each.

As shown in Figure 4, among the analysed credits, MRc1.1 focuses on preserving existing structural elements (walls, floors, and roofs) to minimise CDW. MRc2 aims to improve construction waste management by encouraging material recovery and recycling, and by reducing the amount of waste sent to landfills and incinerators. MRc4, which promotes the use of materials with recycled content, further supports this goal and refers to the ISO 14021 standard for identifying such materials and local waste processors [10]. MRc6 encourages sustainable construction practices by promoting the use of rapidly renewable materials in building projects.

SHAP analysis results show that feature contributions across all four credits can be categorised as high (≥10%), moderate (5–9.9%), low (1–4.9%), or negligible (<1%). This categorisation offers a clear overview of the most influential features for different CDW credit types. The most influential factor across all four credits was material resources, with high contributions of 20.6% (MRc1.1), 53.4% (MRc2), 36.5% (MRc4), and 22.6% (MRc6). Other features with high influence included GER (13.5%) in MRc1.1, GDP per capita (21.5%) in MRc4, regional priority credits (10.8%) in MRc2, and energy and atmosphere (11.1%) in MRc6. Moderate contributions were found for water efficiency (MRc1.1: 9.5%; MRc6: 5.1%), innovation (MRc6: 9.3%), GER (MRc6: 6.8%), and Winter_Temp (MRc1.1: 6.2%). Several features demonstrated a low impact, including energy and atmosphere, innovation, GER, Wind Mps, gross floor area, Temperature Difference, certification level, and regional priority credits, with contributions ranging from 1% to 4.9% across these credits. Predictors with negligible influence (<1%) included firms and building types such as SECH. Although building type is a single categorical factor comprising three binary features (residential, CIO, and SECH), all components showed minimal importance, with SHAP values not exceeding 2.1% for any credit. This indicates that building typology has a limited role in explaining credit success.

Apart from numerical rankings, SHAP analysis provided more profound insights into the context-specific drivers of each credit:

a.

For MRc1.1, where the aim is to preserve structural elements, the impact of material resources, GER, and water efficiency highlights the importance of project-level material planning and environmental performance. Winter temperature and building type (residential) also played moderate roles, while GDP per capita and firm features had a negligible effect.

b.

In MRc2, the prominent role of material resources underscores the central role of effective waste handling strategies. Regional priority credits and indoor environmental quality were also significant, whereas firm-level or typological characteristics had minimal impact.

c.

For MRc4, both material resources and GDP per capita were crucial, highlighting the significance of economic capacity in promoting the use of recycled materials. Features such as innovation, GER, and gross floor area indicate that creativity and project scale also play essential roles.

d.

Finally, MRc6, achieved in only 27 projects, was influenced by material resources, energy, atmosphere, and innovation. Moderate contributions from GER, Temp_Diff, and water efficiency highlight the need for integrated sustainability and climatic adaptability. Once again, firms and regional priority credits had minimal impact.

SHAP-based explainability confirmed that material resources consistently influence CDW credit performance. However, the role of other features, such as environmental efficiency, innovation capacity, and economic context, varies across credit types. In contrast, organisational and structural features (e.g., firm and building type) provide limited explanatory power.

5. Discussion

This study demonstrates the potential of XAI models to enhance transparency in decision-making for CDW credit evaluation, addressing a significant challenge in applying AI to sustainability contexts. AI models often lack interpretability for end-users, a challenge widely recognised in the literature as the ‘black box’ problem [95,96]. In particular, tree-based ensemble models (e.g., XGBoost, CatBoost) are considered black-box models due to the difficulty of interpreting their decision-making processes, despite their high predictive performance [96,97]. SHAP analysis improves transparency by identifying the relative importance of input features [46]. Of the six credits, MRc1.2 and MRc3 were excluded due to insufficient samples; CatBoost was applied to the remaining four, followed by SHAP-based interpretation.

5.1. Interpretation of ML Model Results

Statistical tests confirmed non-normal distributions for MRc1.1, MRc2, MRc4, and MRc6, requiring non-parametric analysis. MRc1.2 and MRc3 were excluded due to their very small sample sizes. SMOTE helped reduce class imbalance but showed mixed results, improving F1 scores in 68.75% of models, lowering them in 27.08%, and leaving the remaining 4.17% unchanged. These results suggest that SMOTE’s effectiveness is context-dependent, varying with model structure and data distribution. Among the 12 machine learning models evaluated, CatBoost consistently delivered the highest performance both before and after applying SMOTE. For MRc1.1, MRc2, MRc4, and MRc6, CatBoost achieved F1 scores of 0.615, 0.944, 0.878, and 0.667, respectively, with corresponding Brier scores of 0.125, 0.088, 0.136, and 0.050. These results demonstrate superior performance in both classification accuracy and probabilistic calibration. Particularly in the MRc6 category, where only 27 projects received points and the highest pre-SMOTE F1 score was 0.364 (achieved by the Random Forest model), CatBoost’s F1 score increased to 0.667 after SMOTE, representing a relative improvement of 83.2%. This finding highlights CatBoost’s robustness in managing imbalanced and complex data structures, underscoring its suitability as a strong candidate for decision-support systems in credit evaluation for CDW.

5.2. Discussion of SHAP-Based Factor Effects

SHAP analysis revealed that Material Resources was the dominant factor across all four credits, contributing 20.6% in MRc1.1, 53.4% in MRc2, 36.5% in MRc4, and 22.6% in MRc6. This highlights the central role of resource efficiency in CDW performance. In MRc2, LEED credit–related features accounted for 80.8% of model decisions, with Material Resources most influential. Similar dominance was observed in MRc4 (60.5%) and MRc1.1 (46.5%). Economic factors, particularly GDP per capita (21.5%), significantly influenced MRc4, while MRc6 exhibited a more balanced structure, shaped by LEED credit features (53.3%), climate (8.9%), and sustainability indicators (8.9%). By contrast, architectural firm features had a negligible influence (≤0.3%), suggesting that the design team’s identity plays a minimal role in CDW credit outcomes.

These results confirm that CDW credit performance is context-specific. Resource efficiency and local material choices, as emphasised in earlier studies [12,21], remain critical drivers, while architectural firms exert minimal influence. CDW credits are thus more strongly shaped by material strategies and project management than by project design. Moreover, higher certification levels improve performance on MRc1.1 and MRc6, whereas MRc2 and MRc4 appear largely independent of certification.

5.2.1. Certification Level

SHAP analyses revealed that certification level had a significant influence on MRc1.1 (3.1%) and MRc6 (2.0%), while having a minimal effect on MRc2 (0.6%) and MRc4 (1.0%). This indicates that MRc1.1 and MRc6 are more closely tied to certification targets, whereas MRc2 and MRc4 are broadly applied regardless of certification level. Dunn’s post hoc tests supported these differences. MRc1.1 showed a significant improvement between Silver and Gold groups (p = 0.0179), suggesting that reuse of existing building components increases with higher certification levels. MRc6 also demonstrated significantly higher success in the Platinum group than in the Gold (p = 0.000452) and Silver (p = 0.000182) groups, highlighting its significant alignment with projects aiming for higher certification. By contrast, MRc2 showed only a non-significant upward trend (p = 0.092), and MRc4 revealed no significant group differences, confirming its wide implementation across certification levels.

5.2.2. Building Type

SHAP analysis reveals that building type has a differential impact on credits. MRc1.1 was the most sensitive (10.9%), with residential projects (6.1%) strongly supporting the reuse of existing components. MRc6 also showed moderate sensitivity (6.3%), with CIO projects (5.1%) adopting a more systematic approach to rapidly renewable materials. In contrast, MRc2 (2.0%) and MRc4 (3.0%) were less affected, indicating that these credits are more dependent on material management and implementation practices than on project type. Among building types, SECH consistently contributed the least (1.3% in MRc1.1; ≤0.2% in MRc2, MRc4, and MRc6). This limited influence likely reflects user-specific requirements and compliance with technical standards, which complicate the reuse of building components [98]. In addition, public accountability and long-term functional sustainability may limit the applicability of environmental performance criteria [5]. These results highlight the need to re-evaluate credit strategies in rating systems like LEED, suggesting that CDW credits should be adapted to project type, particularly for MRc1.1 and MRc6, while MRc2 and MRc4 remain relatively context-independent.

5.2.3. Economic, Sustainability, Climatic Factors

GDP per capita showed varying effects across credits, with the most potent in MRc4 (4.6%) and MRc6 (3.5%). MRc1.1 displayed a moderate impact (1.8%), while MRc2 remained largely unaffected (0.5%). This indicates that economic capacity is more critical for material-focused credits. Project area contributed notably to MRc1.1 (3.3%) and MRc6 (2.7%), but less so to MRc2 (2.1%) and MRc4 (2.0%). The GER feature consistently exceeded 2% across all credits, with the highest values in MRc1.1 (2.9%) and MRc4 (2.8%), suggesting that projects with higher sustainability efficiency are more likely to reuse building components and adopt recycled materials.

Climatic features showed differentiated effects. MRc6 emerged as the most climate-sensitive credit (10.9%), driven by average temperature (2.6%) and sunshine duration (2.2%). MRc1.1 was moderately influenced (7.5%), with temperature (2.3%) and sunshine (1.8%) being the main contributors. By contrast, MRc2 (<1%) and MRc4 (3.4%) showed limited climatic sensitivity, relying more on economic or project-specific strategies. These results emphasise that while economic and sustainability indicators strongly affect MRc1.1, MRc4, and MRc6, climate factors primarily shape MRc6 performance.

SHAP analysis shows that while resource efficiency is universally essential, the roles of certification, project type, economic capacity, and climate vary by credit type, necessitating differentiated, adaptive strategies for LEED credit optimisation.

5.3. Thresholds and Challenges in CDW Credits

These results suggest that contextual and regional factors significantly influence CDW credit preferences. The low success rates in MRc1.2 and MRc3 show that conserving or reusing existing components is particularly difficult in practice. Under LEED-NC 2009, six credits address CDW management. MRc1.1 grants 1–3 points for reusing 55–95% of major components such as walls, floors, and roofs [10,53]. MRc1.2 awards 1 point for reusing 50% of non-structural interior elements, while MRc2 provides 1–2 points for diverting 50–75% of CDW from landfills [9]. Taken together, these thresholds imply an average reuse rate of about 64%, although reuse-focused credits (MRc1.1, MRc1.2, and MRc3) comprise 55% of CDW-related points, MRc1.2 and MRc3 remain scarcely achieved, and even MRc1.1 demonstrates limited success.

High thresholds of 50–55% for MRc1.1 and MRc1.2 are often impractical [16]. In contrast, credits with lower thresholds show more consistent uptake. For example, MRc3 grants points for reusing 5% or 10% of existing structural elements [99], MRc4 requires 10–20% of recycled materials [100], and MRc6 demands 2.5% of rapidly renewable materials [99,100]. Despite similar percentage thresholds, success rates differ markedly. MRc4 has achieved high adoption, while MRc3 remains rarely implemented, and MRc6 shows limited achievement. These differences suggest that threshold design, as much as intent, affects implementation feasibility.

SHAP analysis provides a data-driven basis for improving credit design by identifying contextual drivers linked to underperformance. To strengthen uptake of MRc1.2 and MRc3, incentive mechanisms and clearer policy support are needed. Expanding datasets to include diverse national contexts would help capture the cultural, economic, and regulatory effects on credit achievement, supporting the development of more targeted policy recommendations. Comparative analyses are also essential to evaluate changes introduced in LEED v4 and later versions. Such approaches would make CDW management criteria more adaptive, localised, and practically achievable within GBRS.

6. Conclusions

Selecting appropriate CDW credits in LEED projects involves a complex decision-making process that begins early in the design stage and is influenced by multiple contextual, sustainability, climatic, and economic factors. To enhance the practical relevance of CDW credit analysis, this study compares its model outcomes with related research, emphasising methodological improvements and contextual differences. The findings also deliver actionable insights for construction practitioners and policymakers by linking AI-supported decision tools to sustainable design practices and regulatory frameworks. Importantly, the absence of AI-driven studies specifically addressing CDW credits within GBRS underscores the novelty of this research, which provides a transparent, explainable decision-support framework to fill this critical gap.

This study examined CDW credit preferences in 407 LEED-NC v3 projects from Turkey, Germany, and Spain, utilising 12 machine learning models optimised with Bayes Search and interpreted using SHAP analysis. MRc2 and MRc4 credits achieved higher success rates, while MRc1.1 and MRc6 showed lower success rates. Due to limited data, MRc1.2 and MRc3 were excluded from the analysis. SMOTE improved classification performance on imbalanced datasets, and CatBoost achieved the best performance across the four credits (F1 scores: 0.615–0.944). These findings highlight the importance of tailoring model choice and preprocessing strategies to the characteristics of each credit. SHAP analysis provided transparent explanations of model decisions, showing that Material Resources was the dominant factor (20.6–53.4%), while architectural firm identity had minimal impact. The certification level had a significant influence on MRc1.1 and MRc6 but had little effect on MRc2 and MRc4. The results indicate that regional adaptations and incentive mechanisms are necessary to increase the uptake of low-performing CDW credits, such as MRc1.2 and MRc3, and to align credit implementation with broader sustainability goals. For policymakers, this highlights the need for targeted strategies, while for practitioners, the framework offers early-stage guidance on credit feasibility under different contextual conditions. These insights contribute to a transparent, data-driven decision-support framework that increases interpretability and strengthens project strategies.

This study has limitations. It focuses only on LEED-NC v3 projects from three countries, excludes other rating systems, and does not incorporate sustainability outcomes, user preferences, or construction-specific features. Moreover, only Bayes Search was applied for hyperparameter tuning. Future research should test models developed during the design phase of LEED projects with data from implementation and post-completion stages, integrate additional contextual and sustainability indicators, and compare results across different LEED versions and other GBRS. Such efforts will enhance model comprehensiveness, improve generalisation, and strengthen decision-making for sustainable construction.