Environmental benchmarking of building typologies through BIM-based combinatorial case studies

Automation in Construction 132 (2021) 103980 Available online 1 October 2021 0926-5805/© 2021 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/). Environmental benchmarking of building typologies through BIM-based combinatorial case studies Alejandro Martínez-Rocamora a , * , Carlos Rivera-G´ omez b , Carmen Gal´ an-Marín b , Madelyn Marrero a a ArDiTec Research Group, Departamento de Construcciones Arquitect´ onicas II, Universidad de Sevilla, Av. Reina Mercedes, 4-a, Sevilla, Spain b Departamento de Construcciones Arquitect´ onicas I, Universidad de Sevilla, Av. Reina Mercedes, 2, Sevilla, Spain ARTICLE INFO Keywords: Benchmarking Building information modelling Dwellings Impact categories Life-cycle assessment Machine learning Prediction model Random forest ABSTRACT Integrated life-cycle assessment (LCA) tools have emerged as decision-making support for BIM practitioners during the design stage of sustainable projects. However, differences between methodologies applied for determining the environmental impact of buildings produce significant variations in the results obtained, making them difficult to be compared. In this study, a methodology is defined for generating environmental benchmarks for building typologies through a combination of BIM-based LCA tools and machine learning techniques. When applied to an 11-story residential building typology with 92 dwellings by varying the constructive solutions of façades, partitions, roof and thermal insulation materials, results fall within a range from 360 to 430 kgCO 2 eq/ m 2 . The Random Forest (RF) algorithm is successfully applied for identifying the most decisive variables in the analysis (partitions and façades), and shows signs of being useful for predicting the environmental impact of future constructions and to be applied to the analysis of greater scale urban zones. 1. Introduction The study of the environmental impact of buildings has been a mainstream research topic in recent years, promoted by new international directives for the reduction of energy consumption, CO 2 emissions, and waste generation of this industrial sector [1]. Frequently, 30–40% of the environmental impact of humans on the Earth is attributed to buildings, which has drawn the attention of architects and engineers to the design of more sustainable buildings, with lower energy and resources consumption [2]. Numerous studies have focused on the development of tools and calculation models to determine this environmental impact, whether using single issue indicators, such as the embodied energy, carbon footprint [3] or emergy [4,5], or more complex ones, such as the ecological footprint [6–8], CML or Eco-indicator, also known as multi-variable indicators. However, while appropriate calculation models were being developed to estimate the environmental impact of buildings, technology has made significant advances with the introduction of new tools based on the Building Information Modelling (BIM) methodology [9]. The European Union Public Procurement Directive (Directive 2014/25/EU) [10] states that all the 28 European Member States may encourage, specify or mandate the use of BIM for publicly funded construction and building projects in the European Union by 2016. This has already become mandatory in several countries such as UK, Netherlands, Denmark, Finland and Norway [11]. This fact, along with the trend of the architecture, engineering and construction (AEC) sector toward sustainability make it necessary to provide designers with reliable tools that combine BIM with life-cycle assessment (LCA), which would allow the latter to Abbreviations: A, Air chamber; AC, Aerated concrete; AEC, Architecture, engineering and construction; AI, Acoustic insulation; AP, Acidification potential; BIM, Building Information Modelling; CB, Cored brick; CEC, Constructive Elements Catalogue; CHB, Concrete hollow-core block; CM, Cement mortar; CS, Concrete slab; CT, Ceramic tile; DBL, Default to building life; DHB, Double-hollow brick; DT, Decision tree; EDA, Exploratory data analysis; EIFS, External insulation finish system; EP, Eutrophication potential; EPD, Environmental product declaration; GHG, Greenhouse gas; GP, Gypsum plaster; GW, Glass wool; GWP, Global warming potential; LCA, Life-cycle assessment; LCB, Light-weight clay block; LCI, Life cycle inventory; MAE, Mean absolute error; NRED, Non-renewable energy demand; ODP, Ozone depletion potential; OOB, Out-of-bag; PB, Gypsum plasterboard; PED, Primary energy demand; PUR, Polyurethane rigid foam; RED, Renewable energy demand; RF, Random forest; RMSE, Root-mean-square error; SFP, Smog formation potential; SHB, Single hollow brick; TI, Thermal insulation; TT, Terracotta tile; VB, Vapor barrier; WL, Waterproof layer; XPS, Extruded polystyrene. * Corresponding author. E-mail address: [email protected] (A. Martínez-Rocamora). Contents lists available at ScienceDirect Automation in Construction journal homepage: www.elsevier.com/locate/autcon https://doi.org/10.1016/j.autcon.2021.103980 Received 25 November 2020; Received in revised form 14 September 2021; Accepted 25 September 2021 Automation in Construction 132 (2021) 103980 2 follow the advance that BIM is imposing. BIM environments add a main information layer to building projects such that the designer now draws constructive elements and populates them with characteristics or parameters, which are defined with a level of detail according to the current life cycle stage of the project [12]. Therefore, this technology allows introducing information of different natures, including those data necessary for carrying out an early LCA study of projects from the design stage, which can be useful for decisionmaking before it is actually built [13–16]. Such analysis would allow designing more sustainable buildings by determining with anticipation the environmental impact that these will produce through the consumption of materials and energy, as well as greenhouse gas (GHG) emissions and waste generation [17]. Few developers have worked on this topic, since the implementation of LCA in BIM environments requires programming skills, turning it into a multidisciplinary issue which needs involving environmentalists, architects, engineers, and programmers in order to develop tools that behave properly and be useful for these purposes [18]. One of the main problems detected is the lack of consistency of the results obtained through these tools in comparison with pure LCA tools such as SimaPro, Umberto or GaBi, probably due to the need for simplifying the analysis and make it accessible to professionals in the AEC industry, who are not specialized in environmental impact assessment [19]. Similarly, the inclusion of primary materials as part of a same material or family constitutes a limitation for a detailed LCA [20], unless the life-cycle inventory includes aggregated materials. This problem relates to the difficulties in defining a framework of necessary parameters to carry out a correct LCA in BIM for the various stages of the life cycle of buildings, which has already been approached in some studies [21,22]. Among the scarce tools currently available, a set of external applications are able to import and work with data from BIM files (Elodie, eTool LCA, GBAT, Impact Compliant Suite, and Lesosai), and few examples of tools have been completely integrated in BIM environments through plugins, such as LCA Design, Ecotect, Green Building Studio, One Click LCA, and Tally [19]. BIM-LCA integration tools allow automating the connection between the information contained in the BIM model, concretely materials quantity take-off, and LCA databases, thus considerably accelerating the environmental impact assessment of architectural projects in early stages. However, a state of the art and a reference baseline are required in order to provide practitioners not only with results, but also with benchmarks that allow them to evaluate if their projects are sustainable or not [23]. Moreover, the aforementioned integration makes it possible to automate the generation and evaluation of multiple project's design variations or alternatives by modifying the necessary parameters within the BIM model. The analysis of a significant quantity of variations generated for an architectural project can be useful to provide the required environmental benchmarks for its specific building typology. It is worth noticing that the analysis of results from these variations can become unmanageable and not very intuitive through simple observation given the great amount of data generated. Machine learning algorithms can be of help to carry out this analysis and provide objective answers and explanations for the results obtained. In addition, machine learning models trained with this data can make better predictions of the environmental impact of buildings. The present study aims at defining a methodology for generating environmental benchmarks for building typologies using a combination of BIM-based LCA tools and machine learning algorithms. The methodology is subsequently tested by means of an actual example of residential building typology as case study for which a set of environmental impact indicators derived from combinatorial variations of its constructive solutions are calculated. These benchmarks will be useful in the short term for promoting a continuous improvement of the sustainability of buildings. In the next section, a state-of-the-art on BIM-LCA integration and environmental benchmarking for buildings is provided. In Section 3, a case study of an 11-story building is presented as a basis for the subsequent combinatorial analysis, along with the research methodology that will be applied in this study. This methodology consists of working with the case study and data in several environments, concretely comprising Autodesk Revit as BIM software for modelling, Tally as integrated LCA plugin, and finally Microsoft Excel for spreadsheets and Python's Scikit- Learn library for data analysis. In Section 4, the results obtained from applying the described methodology to the case study are analyzed through a combination of exploratory data analysis and machine learning techniques. In Section 5, these results are discussed in comparison to those from similar studies. Finally, in Section 6, the conclusions drawn from this study are presented. 2. State-of-the-art 2.1. BIM-LCA integration A general overview of BIM-based tools for environmental impact assessment showed that most of them need to combine BIM software with other applications to obtain quantification of environmental indicators. Various approaches have been followed, which were classified by Santos et al. [22]: using external tools [24]; connecting quantity takeoff and an external LCA database with integrated tools in proprietary software [25]; and including LCA information in BIM models [22]. The main advantage of this last approach is that it automatically updates results when the project is modified, thus taking advantage of the full potential of BIM tools, while the others require to re-export the information and re-link it to the external databases, besides needing extra licenses for the external software involved. Regarding the first approach, Jrade and Jalaei [26] developed an integration between Autodesk Revit, Microsoft Excel, and Athena Impact Estimator to analyze the environmental impact and the EE of constructive solutions. Marzouk et al. [27] proposed a combination of several software tools (i.e., Autodesk Revit, Revit DB Link, Microsoft Access, and Athena Impact Estimator) for measuring direct and indirect CE in construction projects. Chen and Pan [28] presented a multicriteria decision making on low carbon construction measures by combining Revit, eQuest, and Promethee [29]. Ajayi et al. [30] explored a combination of Revit Architecture, Green Building Studio (GBS), and ATHENA complement for global warming potential (GWP) and health impact assessment. Crippa et al. [31] integrated data extracted from Ecoinvent with Excel and ArchiCad for the analysis of the carbon footprint of wall systems in Brazil. Finally, the authors also suggest the development of an automation module (a plugin) within the BIM environment through its application programming interface. However, these integration models frequently require some manual steps to be taken by the user, which is somehow opposite to the pursued automation of the analysis. As examples of application of the second approach, Najjar et al. [32] analyzed a case study with the Tally plugin to evaluate the entire life cycle of buildings, aiming to provide recommendations such as carefully reviewing the manufacturing process and technological advances of the construction materials selected for an architectural project in order to reduce the environmental impact of both the manufacturing and operation phases. Schultz et al. [33] studied the differences in the LCA results obtained from Tally and Athena Impact Estimator, those of general use in USA, getting similar results, but identifying significant differences in the impact categories they presented. Basbagill et al. [24] developed a BIM model using DProfiler linked to eQuest within a BIM environment. Subsequently, the results were manually loaded into SimaPro and Athena EcoCalculator to obtain the CF. Azhar et al. [34] combined Revit with IES Virtual Environment to calculate CE and incorporated energy to convert them into LEED credits. Also, Ilhan and Yaman [18] developed a green building evaluation tool using Graphisoft ArchiCAD® linked to the BREEAM material database. Regarding the third approach, some authors reviewed the existing A. Martínez-Rocamora et al. Automation in Construction 132 (2021) 103980 3 research on BIM-LCA integration and its possibilities for simplification in terms of input-output data and LCA results [14], and pointed out that the best solution would be to remain within the BIM environment to facilitate the interaction between design and environmental performance assessment [2]. Gan et al. [35] developed a holistic approach to evaluate built-in and operational carbon in high-rise buildings using Revit and a parametric complement, Dynamo. These researchers also evaluated the reduction of the CF by replacing steel and cement with recycled materials [36]. Other environmental evaluation tools and models were defined by Inyim et al. [37], who presented a BIM extension named Simulation of Environmental Impact of Construction (SimulEICon), which was designed to assist in decision-making during the design stage of construction projects. Finally, Yang et al. [38] deployed a unidirectional workflow to calculate the operation and energy of materials and CO 2 eq. 2.2. Environmental benchmarking of buildings As mentioned before, some researchers highlight the need for benchmarks that provide designers with references to evaluate the environmental performance of their projects [23]. These benchmarks would also allow the establishment of policies with sustainability requirements for new buildings [39]. Recently, Trigaux et al. [40] analyzed 23 different systems to generate environmental benchmarks based on five categories comprising 16 sub-aspects that can vary between systems, which produced different results depending on the system used. For example, Alghamdi et al. [41] used fuzzy clustering to group educational building typologies by energy and water consumption and CO 2 emissions. This allowed them to identify those needing more urgent refurbishment in order to reduce the environmental impact of the built environment. Ji et al. [42] established baselines for the environmental impact of educational buildings referred to 8 indicators and economic cost. To that end, case studies where classified into two clusters, obtaining carbon footprint reference values of 2530 and 3700 kgCO 2 eq/m 2 for them. Moschetti et al. [43] analyzed four residential building typologies introducing variations in their constructive solutions based on the results from the European project TABULA, and locating them in three climatic zones in Italy aiming at establishing reference values in terms of economic cost and environmental impact indicators. As can be seen, although some initial approaches to this topic have been carried out by researchers, environmental benchmarking of buildings remains a scarcely explored research field, and robust methodologies to establish these benchmarks must arise to provide practitioners and policy-makers with reference baselines. The present study defines a method to automate the benchmarking of building typologies through a sequence of generation of combinatorial variations of constructive solutions in BIM, quantity take-off and LCA, followed by a data selection and sorting process based on spreadsheet macros, and finally data analysis with standard machine learning algorithms. This method is not only capable of providing an environmental benchmark for any building typology, but also allows identifying the most influencing variables in order to support efficient decision-making during the design process. 3. Materials and methods 3.1. Description of the case study The selected case study, located in the north sector of Seville (Spain, 37◦22′58” N 5◦58′23” W, 16 m.a.s.l.), is an 11-story residential building with 92 dwellings, commercial premises in the ground floor, and two underground floors containing the parking lot and storage rooms. It is a detached building with 13,367.93 m 2 of built-up surface, which was built in 2002 and, for the purposes of this study, a service life of 100 years is assumed (according to Spanish Regulations [44]). It has an almost rectangular shape with two small wings, and a front yard where pedestrian and vehicular accesses are located (see Fig. 1). The building's typology is a representative case of social dwellings built in Spain in the latest decades [45–47], which allows results to be extrapolated to a significant amount of existing buildings in that region. The building's structure consists of a 70 cm-thick reinforced concrete slab foundation, retaining walls in the underground levels, and columns and waffle slabs in the rest of floors. There are three staircases and four elevators distributed in three main structurally independent blocks. Each block contains four dwellings per floor, except for the last two floors, where four dwellings in total share the space with an accessible rooftop. The original envelope is built with 9 cm-thick masonry walls. This brick façade system offers mostly a facing brick finish and cement mortar lining in some areas, with polyurethane foam insulation plus an air chamber and self-supporting plasterboard partitions in the inner side. The floorings are generally covered with terrazzo, and ceilings with gypsum plasterboard. In the kitchens and bathrooms, ceramic tiles fully cover walls and floors. The underground floors use cement mortar-based covering in floors. Given that technical services will not suffer variations in this study, in addition to not being included in Tally, these are not taken into consideration in the system boundaries, and therefore do not need to be described. 3.2. Research methodology In order to ease understanding of the following sections, an overview of the research methodology applied in the present study is provided in Fig. 2, where the various steps to be taken and the environments or tools with which these are carried out are specified. As a first step, it is necessary to make a selection of the constructive solutions to be included in the study as variations for the building elements. Then, types of BIM elements are generated for each constructive solution, and entries from the LCA database are assigned to the materials involved in these solutions, thus defining what is called a template file. Once the template file is ready, the building typology is modeled, and BIM types can be assigned to the elements in the model in order to generate all the different combinations that will become case studies. Subsequently, the various combinations are analyzed with the LCA plugin, generating a report file for each case. The information in these reports is then automatically structured as desired through macros that read the source files and organize the data in a new spreadsheet. In the next step, exploratory data analysis (EDA) is applied to these data through Excel and Python's Scikit-Learn library in order to identify outliers, correlations, and the impact of each variable in the results obtained. Finally, some of the findings from EDA are double-checked by using a Random Forest (RF) regression algorithm, which aims to predict results for future combinations or case studies that match this building typology. 3.2.1. LCA software, database, and assessment methodology For the purposes of this study, the Tally plugin for Autodesk Revit (ver. 2019.06.27.01) [48] has been selected as the tool for the environmental impact assessment of the building. This plugin uses the GaBi database as LCA data source, which was identified as one of the most complete and comprehensive databases for LCA of construction materials [49]. First, it scans the BIM model and identifies all the elements and materials in the project. Subsequently, entries from the LCA database must be assigned to each material. Parameters such as density, service life or transportation means and distances must be specified to better describe them. After the analysis is performed, several indicators are reported, including materials mass, potentials of acidification, eutrophication, global warming, ozone depletion and smog formation, and primary, non-renewable and renewable energy demands. The presentation and analysis of results in this study considers all these indicators. However, for a richer comparison to other studies, the discussion section is focused A. Martínez-Rocamora et al. Automation in Construction 132 (2021) 103980 4 on the carbon footprint indicator, that is, global warming potential, which consists of determining the emissions of GHG produced by a specific process [50]. While other indicators are of great importance, carbon footprint stands out for its strong relation to the main aims of the Kyoto Protocol, and the ease to communicate results to non-specialized audiences [51]. Regarding the system boundaries, all the life-cycle stages of construction materials, A1 to C4 and D according to UNE-EN 15978:2012 [52] will be considered in this analysis, with the use stage being blocked for all the cases product of the combinatorial assessment by adjusting the thermal insulation thickness in order to produce equal thermal transmittances for each one of the selected envelope solutions. Therefore, only maintenance (B2) and replacement (B4) of materials are included regarding stage B of the life cycle. While the building's structure is not affected by these variations of constructive solutions, it has been included in the analysis since it usually produces a significant share of the total environmental impact of building materials, and the influence of the variations here proposed in the overall result of the building is worth being studied. As mentioned before, Tally still does not consider technical services in its analysis, and therefore these cannot be included in the assessment. Fig. 1. Case study of a 10-story residential building in Seville. A. Martínez-Rocamora et al. Automation in Construction 132 (2021) 103980 5 3.2.2. Selection of constructive solutions and assignment of LCA data entries In order to generate populated distributions with the results obtained, a selection of constructive solutions was made that will be considered for the multiple variations of the case study. These solutions were obtained from the Constructive Elements Catalogue (CEC) of the Technical Code of Buildings Construction in Spain [53], which contains a comprehensive list of solutions for façades, partitions and rooftops. As shown in Table 1, this selection considers 13 different façade solutions, including options with external insulation finish systems (EIFS), 4 types of internal partitions, and 3 rooftop solutions which are the most widespread constructive solutions in Spain [45–47,54]. Façades, partitions and rooftop and their finishes have been identified as decisive constructive elements for the environmental impact of buildings [55]. Additionally, variations of these are considered regarding the type of thermal insulation material used and the need for ceramic tile coverings for wet rooms in one or both sides, according to their location in the building. Thus, according to the solutions presented in this table, it will be possible to generate 240 (39 ×4 ×5) different versions of the case study by modifying the family types assigned to the façades, partitions and rooftops in the model, with side walls being modified according to the solution applied to partitions. To that end, it is necessary to first generate a template file containing the family types and the assignment of their materials to specific entries of the LCA database. Given that this study aims to provide a methodology for generating environmental benchmarks for building typologies, this pairing between BIM materials and LCA database entries becomes crucial for the traceability and reproducibility of results by future researchers, since different methodologies and pairings could produce significantly different results [56]. Therefore, in Appendix 1, Table A.1 shows the assignment of LCA database entries to the materials of the various constructive solutions considered in the combinatorial case study, as well as extra parameters specified for each assignment, and Table A.2 presents the transportation distances by truck applied to each material considering the case study's location. 3.2.3. Data processing and analysis The results from each generated design option will be obtained in individual spreadsheets where the contribution of each item to the total environmental impact is specified by life cycle stage, division (e.g. concrete, masonry, openings and glazing, finishes, etc.), and category (e. g. doors, floors, roofs, walls, structure, etc.), thus allowing a detailed analysis of the source for each impact. Given the great amount of combinations (240), and therefore result files, along with the fact that the corresponding BIM models will have to be manually generated, the processing of results will be automated through a macro which will read every file in the folder and will combine the data in a single spreadsheet in order to allow a comparative analysis. Finally, an exploratory data analysis (EDA) methodology will be applied for analyzing the distributions of overall results for the 240 cases regarding each environmental impact indicator, to subsequently search for the specific causes for the identified differences. Additionally, a massive analysis of combinations will allow determining reference ranges of results for this building typology as a baseline for the evaluation of the sustainability of design options for new buildings. The analysis will be supported by the design of a RF regression model which, among other functionalities, allows identifying the most important features that determine the results. 3.2.4. Random forest algorithm As mentioned before, while the application of EDA to the results obtained can be descriptive for the 240 variations generated in this study, this could be less intuitive for a greater set of combinations that involved variations in windows, doors, floor slabs, floorings, paints, and technical services elements, among others. Machine learning algorithms can provide a consistent analysis for a massive amount of data and help in obtaining objective findings based on statistics. Decision trees (DT) are tree-based machine learning algorithms that allow executing both classification and regression tasks. In decision trees, every node is a condition on how to split values in a single feature so that similar values of the dependent variable end up in the same set after the split [57]. The features for internal nodes are selected with some criterion, which for classification tasks can be Gini impurity or information gain, and for regression tasks is variance reduction [58]. DTs allow computing how much each feature contributes to decreasing the weighted impurity or, in the case of regression tasks, to reducing variance. In other terms, this reduction of variance indicates which features are more decisive for obtaining the final result in the Fig. 2. Methodology flowchart. A. Martínez-Rocamora et al. Automation in Construction 132 (2021) 103980 6 regression task and thus, in our case, which features can produce greater reductions in the building's environmental impact. In the case of RFs, the importance of features is measured by averaging the reduction in variance over trees [57]. The relative importance of features can also be obtained for other machine learning algorithms such as neural networks by calculating permutation importance, although this requires further manipulation of the dataset [59]. RF automatically provides an easily accessible ‘feature importance’ variable that is calculated after training, which makes this machine learning algorithm especially useful for the present study's purposes. RF is one of the most frequently used machine learning algorithms nowadays, and it has been successfully applied before in classification and regression tasks related to the construction sector [60–62]. It consists of an ensemble of DTs, each one generated with a random subset of the same training set. These subsets can be obtained whether applying bootstrap sampling, i.e. sampled with replacement, or not. As an ensemble method, its results are subsequently determined by aggregating the results obtained in the various DTs as a mean value. If bootstrap sampling is applied, the RF behaves as a bagging (bootstrap aggregating) algorithm. Otherwise it is called a pasting algorithm [63]. Bagging implies that, probabilistically speaking, around one third of the samples in the training set will not be sampled in each decision tree, being these known as ‘out-of-bag’ (OOB) samples. In an ideal case, about 36.8% of the total training data forms the OOB sample in each tree [64], but it is worth noting that this subset of samples will not be identical for all the decision trees in the RF. OOB samples allow determining the socalled OOB score, which is the mean coefficient of determination (R 2 ) for all the OOB samples, which are evaluated in the DTs that did not sample them for training. This means only a subset of DTs can be used for determining the coefficient of determination for each OOB sample. In general, validation on a full ensemble of DTs is better than a subset of DTs for estimating the score. However, occasionally the dataset is not big enough, and hence setting aside a part of it for validation is unaffordable. Consequently, in cases where a large dataset is not available and it must be all consumed as the training dataset, the OOB score provides a good trade-off [64] and insights on how the model could behave with a larger dataset that allowed splitting into training and validation sets. 4. Results 4.1. Overall environmental impact of combinatorial case studies As a first approach to analyze the results of the environmental impact assessment performed with Tally, the scatter plots of the total results obtained for the various ecological indicators are presented with the case number, i.e. combination of constructive solutions, in the horizontal axis (Fig. 3). These results show an apparent similarity among them, which can indicate a strong correlation, except for those regarding Ozone Depletion Potential (ODP). This is double-checked through a correlation matrix generated with Python's numpy, seaborn and matplotlib libraries (Fig. 4), where all indicators except ODP show values close to 1, thus confirming that strong correlation between them. As can be observed, combinations in the first half (1−120) produce higher environmental impacts than those of the second half (121–240). By analyzing the variations introduced as input variables between these two subsets of combinations, the cause of this difference in results seems to be the constructive solution assigned to partitions. Cases 1 to 120 use P1.1 and P2.1 from the solutions described in Table 1 (represented as blue points in Fig. 3), that is, single hollow bricks covered with gypsum plaster and paint, while cases 121 to 240 use P4.1 and P4.2 from Table 1, i.e. plasterboard-based solutions (represented as orange points in Fig. 3). Another observable difference in results is detected between each set of 15 combinations or cases, which are highlighted with different gradients of colors in Fig. 3. The variation introduced in these subsets corresponds to the façade solution, with higher environmental impacts Table 1 Selection of constructive solutions for a combinatorial case study. Code Base layers, outer to inner (thickness (cm)) Thermal insulation Façades Material Thickness (cm) F1.1 CB (11.5) +CM (1.5) +TI +DHB (7.0) + GP (1.5) Glass fiber 2.0 PUR foam 1.8 EPS 2.1 F1.3 CB (11.5) +CM (1.5) +TI +PB (1.1) Glass fiber 2.4 PUR foam 2.2 EPS 2.6 F1.9 CHB (14.0) +CM (1.5) +TI +DHB (7.0) +GP (1.5) Glass fiber 2.1 PUR foam 1.9 EPS 2.3 F1.12 CHB (14.0) +CM (1.5) +TI +PB (1.1) Glass fiber 2.5 PUR foam 2.3 EPS 2.8 F3.1 CM (1.5) +CB (11.5) +TI +DHB (7.0) + GP (1.5) Glass fiber 2.0 PUR foam 1.8 EPS 2.1 F3.3 CM (1.5) +CB (11.5) +TI +PB (1.1) Glass fiber 2.4 PUR foam 2.2 EPS 2.6 F3.9 CM (1.5) +CHB (14.0) +TI +DHB (7.0) +GP (1.5) Glass fiber 2.1 PUR foam 1.9 EPS 2.3 F3.13 CM (1.5) +CHB (14.0) +TI +PB (1.1) Glass fiber 2.5 PUR foam 2.3 EPS 2.8 F3.21 CM (1.5) +LCB (14.0) +TI +DHB (7.0) +GP (1.5) Glass fiber 1.7 PUR foam 1.5 EPS 1.8 F3.23 CM (1.5) +LCB (14.0) +TI +PB (1.1) Glass fiber 2.1 PUR foam 1.9 EPS 2.3 F4.1 CM (1.5) +TI +CB (11.5) +GP (1.5) Glass fiber 2.5 PUR foam 2.3 XPS 2.8 F4.3 CM (1.5) +TI +CHB (14.0) +GP (1.5) Glass fiber 2.5 PUR foam 2.3 XPS 2.8 F4.5 CM (1.5) +TI +LCB (14.0) +GP (1.5) Glass fiber 2.1 PUR foam 1.9 XPS 2.3 Partitions P1.1 GP (1.5) +SHB (4.0) +GP (1.5) – – P2.1 GP (1.5) +SHB (4.0) +AI +SHB (4.0) + GP (1.5) Mineral wool 3.0 P4.1 PB (1.1) +A +PB (1.1) – – P4.2 PB (1.1) +PB (1.1) +AI +PB (1.1) +PB (1.1) Mineral wool 3.0 Rooftops C1.6 TT (1.2) +CM (1.5) +WL +TI +VB + AC +CS (30.0) PUR foam 5.7 XPS 6.9 Glass wool 6.3 C1.6i TT (1.2) +CM (1.5) +TI +WL +AC +CS (30.0) XPS 6.9 C4.6 CT (1.2) +A +CM (1.5) +TI +WL +AC +CS (30.0) XPS 6.9 Side walls P2.3 GP (1.5) +CB (11.5) +TI +CB (11.5) + GP (1.5) EPS 3.0 P2.5 GP (1.5) +CHB (14.0) +TI +CHB (14.0) +GP (1.5) EPS 3.0 P2.4 GP (1.5) +LCB (14.0) +TI +LCB (14.0) +GP (1.5) EPS 3.0 CB: Cored brick; DHB: Double hollow brick; SHB: Single hollow brick; CHB: Concrete hollow-core block; LCB: Light-weight clay block; CM: Cement mortar; GP: Gypsum plaster; PB: Gypsum plasterboard; TI: Thermal insulation; AI: Acoustic insulation; A: Air chamber; WL: Waterproof layer; VB: Vapor barrier; TT: Terracotta tile; CT: Ceramic tile; AC: Aerated concrete; CS: Concrete slab. A. Martínez-Rocamora et al. Automation in Construction 132 (2021) 103980 7 in cases 1–15, 31–45 and 61–75, while cases 16–30, 46–60 and 91–105 show lower results in the first half. Those with a higher impact have in common the use of ceramic bricks in both the outer and inner layers, while the latter use concrete hollow-core blocks (CHB) in one of their layers. Concretely, cases 31–45, which have cored bricks in the outer layer and double hollow bricks in the inner layer, the most common configuration in the project location, produce the highest environmental impact of all the analyzed combinations and of those in each half, and cases 91–105, with an EIFS solution as outer layer and CHB as inner layer, show the lowest impact in this first half of combinations. Regarding the second half, that is, those using P4.1 and P4.2 as partitions, those variations of the façade solution with higher Fig. 3. Environmental impact of the 240 combinations for the case study expressed in terms of the ecological indicators included in Tally. A. Martínez-Rocamora et al. Automation in Construction 132 (2021) 103980 8 environmental impact coincide with those in the first half. However, the combinations causing the lowest impacts differ. Specifically, cases 136–150, 166–180 and 211–225 again share the use of CHB as support layer instead of ceramic bricks. This difference in the lowest impacting solutions between the first and the second half of combinations is due to the fact that, in the configuration of cases, when plasterboard partitions are assigned, plasterboard inner layers are also assigned to façades as in fact occurs in construction, thus generating a lower environmental impact than brick-based inner layers. Consequently, cases 136–150, i.e. those using P4.1 and P4.2 for partitions and F1.12 for façades, cause the lowest environmental impact of all the analyzed combinations. As mentioned before, variations in ODP results follow a different pattern, with those cases using glass wool (GW) as thermal insulation material in roofs showing the highest impact in this category. According to consulted Environmental Product Declarations (EPDs) of GW [65], expanded polystyrene (XPS) [66] and polyurethane rigid foam (PUR) [67], PUR should produce a higher ODP than XPS and GW due to the use of CFC-11 gas as inflating agent, which points out the possibility of a mistake in the calculations of this version of Tally for that indicator. For that reason, it is decided to neglect the ODP impact category in this study. Apparently, the partitions and façade solutions are the most decisive variables in defining the building's environmental impact since they cause the greatest differences. This might be somehow easy to perceive from the scatter plots in Fig. 3. However, if there were a greater quantity of variables involved in the design process decision-making, as there should be in order to determine environmental benchmarks for building typologies, this analysis would become less viable. To that end, in the next subsection a regression model based on the RF algorithm is applied with two main aims: first, to corroborate the aforementioned findings about the importance of each variable in determining the final results; and second, to prove the usefulness and feasibility of the methodology here described to analyze and predict the environmental impact of a larger set of building typologies by using artificial intelligence techniques such as RF. 4.2. Random forest regression As mentioned in subsection 3.2.4, when applying RF to design a prediction model with a reduced dataset, setting aside a part of it for validation becomes unaffordable. In these cases, it is usually preferred to use the entire dataset as training set and use OOB evaluation to validate the model, bearing in mind that OOB evaluation is much less reliable than using a validation set [64]. The RF model was trained using five features as input variables, corresponding to the constructive solutions specified in Table 1 for: façade, façade's thermal insulation material, partitions, roof, and roof's thermal insulation material. The output variables were the 8 indicators, once ODP was discarded. The number of estimators, i.e. DTs, generated during the training process was determined by studying configurations from 1 to 50 trees for each environmental impact indicator. Bootstrap and oob_score parameters were set to true. These configurations were evaluated using three of the most usual statistical parameters, the coefficient of determination (R 2 ), the mean absolute error (MAE), and the root mean squared error (RMSE) [63]. Additionally, the potential of the model to make accurate estimations with new data was taken into account based on the OOB score. Despite Fig. 5 only shows this evaluation for the GWP indicator, given the strong correlation between them (see Fig. 4), the result obtained was the same for all the output variables. As can be seen, R 2 soon reaches values close to 1, and the OOB score starts from 0.998 with 10 trees and then slightly increases. MAE and RMSE are soon reduced to less than 2000 and 3500, respectively, reaching a minimum with 43 trees. At this point, also a maximum is reached by the OOB score, thus making it the best option among the evaluated variations. Considering that results for GWP in the dataset vary from 4.8⋅10 6 to 5.8⋅10 6 kgCO 2 eq, 2000 kgCO 2 eq represent an error of 0.00038% in the estimations. It is worth bearing in mind that this error rate corresponds to a validation based on the OOB-score and not a separate validation set evaluated on all the DTs in the RF, which could be higher. In future developments of this research, including a wider set of variations, the number of observations would allow splitting the dataset into training and validation sets, thereby obtaining more Fig. 4. Correlation matrix of output variables in the study. A. Martínez-Rocamora et al. Automation in Construction 132 (2021) 103980 9 accurate evaluation results. Once the number of trees was decided, the RF algorithm was applied to each output variable to analyze the importance of each feature or input variable for determining the estimations. As shown in Table 2, the constructive solutions for partitions and façades are the two most important variables for all the environmental impact indicators, thus confirming the findings from the previous subsection. Also the OOB score obtained for each output variable provides insights on the performance that the RF models could have on new data, thereby proving the feasibility of defining environmental impact prediction models using a set of input variables that can be controlled within the BIM environment. 4.3. Detailed environmental impact When studying the environmental impact according to the three classification modalities Tally has to offer, now focusing on GWP (see Fig. 6), it can be seen that the product and maintenance stages account for a high portion of the emissions, with ~3000–4000 and ~ 1750 tCO 2 eq, respectively. In combinations using brick walls as partitions and inner layers of façades, concrete (~1800 tCO 2 eq), masonry (~1000 tCO 2 eq) and finishes (~860 tCO 2 eq) account for 96% of the GWP during the product stage, which is logical as these products represent most of the mass of construction materials in the building and, as stated before, GWP (and the rest of indicators) has a strong correlation with mass. In cases where the partitions and the façade's inner layer is solved with plasterboard, masonry is relegated to third place with ~430 tCO 2 eq, but still the sum of concrete, masonry and finishes account for 93% of the product stage GWP. Regarding the maintenance and replacement stage, finishes of floors, walls and ceilings, along with doors and windows, produce 94.5% of this stage's GWP, which is logical given that these elements have shorter service lives (see Table A.1) and therefore must be replaced at least once during the building's service life. In contrast, the difference between the GWP of brick walls and plasterboard during this stage is not significant. When observing the results by division, it is observed that the building's overall GWP is mainly caused by finishes (~2400 tCO 2 eq), concrete (~1800 tCO 2 eq) and masonry (~150–1250 tCO 2 eq), with 97% of the former being explained by the use of terrazzo and ceramic tiles for flooring, plasterboard for ceilings, and cement stucco for exterior finishes, while the importance of concrete and masonry was expected given their high presence in the building's structure and walls. Again, the high variability of the masonry's GWP is due to changes in the materials used for partitions and the inner layer of façades, obtaining lower environmental impacts when using plasterboard instead of bricks. Regarding GWP by Revit category, Tally provides slightly misleading results, since by default it includes the structural layer of floors, roofs and walls within these categories instead of separating and including them in the structure category, where only columns are originally considered. Therefore, a corrected version of this last graph was Fig. 5. Statistical parameters in the RF regression model training to estimate GWP with variations in the number of trees. Table 2 Feature importance and statistical parameters of each input variable for the prediction of the environmental impact indicators using RF regression with 43 trees. Environmental impact indicator AP EP GWP SFP PED NRED RED Mass Feature Importance (%) Partitions 49.76 59.50 60.69 55.44 58.27 57.54 68.20 78.84 Façade 46.41 39.32 37.99 43.06 41.50 42.23 31.44 21.13 TI Façade 1.40 0.40 1.09 0.41 0.17 0.18 0.15 0.00 Roof 1.31 0.45 0.19 0.79 0.05 0.05 0.09 0.03 TI Roof 1.12 0.33 0.04 0.30 0.01 0.00 0.12 0.00 Statistical parameters Mean 21,667.78 1435.98 5.29⋅10 6 3.58⋅10 5 7.47⋅10 7 6.58⋅10 7 8.85⋅10 6 1.46⋅10 7 SD 480.50 29.31 2.83⋅10 5 8508.10 5.04⋅10 6 4.72⋅10 6 3.20⋅10 5 1.02⋅10 6 R 2 0.9990 0.9997 0.9999 0.9995 0.9999 0.9999 0.9999 1.0000 OOB 0.9928 0.9975 0.9991 0.9966 0.9995 0.9995 0.9992 1.0000 MAE 9.83 0.41 1937.96 135.53 30,772.45 26,910.65 2525.31 898.52 RMSE 14.78 0.54 3235.19 184.78 41,966.87 38,443.10 3441.80 1452.25 A. Martínez-Rocamora et al. Automation in Construction 132 (2021) 103980 16 Energy Build. 155 (2017) 339–351, https://doi.org/10.1016/j. enbuild.2017.09.038. [9] M. Mousa, X. Luo, B. McCabe, Utilizing BIM and carbon estimating methods for meaningful data representation, in: Procedia Engineering, Elsevier Ltd, 2016, pp. 1242–1249, https://doi.org/10.1016/j.proeng.2016.04.160. [10] European Union, Directive 2014/24/EU of the European Parliament and of the Council of 26 February 2014 on Public Procurement and Repealing Directive 2004/18/EC Text with EEA Relevance. https://eur-lex.europa.eu/legal-content/ EN/TXT/?uri=celex:32014L0024, 2014. [11] European Commission, European Construction Sector Observatory: Building Information Modelling in the EU Construction Sector. https://ec.europa.eu/doc sroom/documents/34518/attachments/1/translations/en/renditions/native, 2019. [12] F.K.T. Cheung, J. Rihan, J. Tah, D. Duce, E. Kurul, Early stage multi-level cost estimation for schematic BIM models, Autom. Constr. 27 (2012) 67–77, https:// doi.org/10.1016/j.autcon.2012.05.008. [13] J.K.W. Wong, J. Zhou, Enhancing environmental sustainability over building life cycles through green BIM: a review, Autom. Constr. 57 (2015) 156–165, https:// doi.org/10.1016/j.autcon.2015.06.003. [14] B. Soust-Verdaguer, C. Llatas, A. García-Martínez, Critical review of bim-based LCA method to buildings, Energy Build. 136 (2017) 110–120, https://doi.org/10.1016/ j.enbuild.2016.12.009. [15] S. Eleftheriadis, D. Mumovic, P. Greening, Life cycle energy efficiency in building structures: a review of current developments and future outlooks based on BIM capabilities, Renew. Sust. Energ. Rev. 67 (2017) 811–825, https://doi.org/ 10.1016/j.rser.2016.09.028. [16] T. Tan, G. Mills, E. Papadonikolaki, Z. Liu, Combining multi-criteria decision making (MCDM) methods with building information modelling (BIM): a review, Autom. Constr. 121 (2021), https://doi.org/10.1016/j.autcon.2020.103451. [17] T. Kulahcioglu, J. Dang, C. Toklu, A 3D analyzer for BIM-enabled life cycle assessment of the whole process of construction, HVAC&R Res. 18 (2012) 283–293, https://doi.org/10.1080/10789669.2012.634264. [18] B. Ilhan, H. Yaman, Green building assessment tool (GBAT) for integrated BIM- based design decisions, Autom. Constr. 70 (2016) 26–37, https://doi.org/10.1016/ j.autcon.2016.05.001. [19] C. Bueno, M.M. Fabricio, Comparative analysis between a complete LCA study and results from a BIM-LCA plug-in, Autom. Constr. 90 (2018) 188–200, https://doi. org/10.1016/j.autcon.2018.02.028. [20] A. Stadel, J. Eboli, A. Ryberg, J. Mitchell, S. Spatari, Intelligent sustainable design: integration of carbon accounting and building information modeling, J. Prof. Issues Eng. Educ. Pract. 137 (2011) 51–54, https://doi.org/10.1061/(ASCE)EI.1943- 5541.0000053. [21] C. Cavalliere, G.R. Dell’Osso, A. Pierucci, F. Iannone, Life cycle assessment data structure for building information modelling, J. Clean. Prod. 199 (2018) 193–204, https://doi.org/10.1016/j.jclepro.2018.07.149. [22] R. Santos, A.A. Costa, J.D. Silvestre, L. Pyl, Integration of LCA and LCC analysis within a BIM-based environment, Autom. Constr. 103 (2019) 127–149, https://doi. org/10.1016/j.autcon.2019.02.011. [23] A. Hollberg, T. Lützkendorf, G. Habert, Top-down or bottom-up? – how environmental benchmarks can support the design process, Build. Environ. 153 (2019) 148–157, https://doi.org/10.1016/j.buildenv.2019.02.026. [24] J. Basbagill, F. Flager, M. Lepech, M. Fischer, Application of life-cycle assessment to early stage building design for reduced embodied environmental impacts, Build. Environ. 60 (2013) 81–92, https://doi.org/10.1016/j.buildenv.2012.11.009. [25] H. Kreiner, A. Passer, H. Wallbaum, A new systemic approach to improve the sustainability performance of office buildings in the early design stage, Energy Build. 109 (2015) 385–396, https://doi.org/10.1016/j.enbuild.2015.09.040. [26] A. Jrade, F. Jalaei, Integrating building information modeling (BIM) and energy analysis tools with green building certification system to conceptually design sustainable buildings, Build. Simul. 6 (2013) 429–444, https://doi.org/10.1007/ s12273-013-0120-0. [27] M. Marzouk, E.M. Abdelkader, K. Al-Gahtani, Building information modelingbased model for calculating direct and indirect emissions in construction projects, J. Clean. Prod. 152 (2017) 351–363, https://doi.org/10.1016/j. jclepro.2017.03.138. [28] L. Chen, W. Pan, A BIM-integrated fuzzy multi-criteria decision making model for selecting low-carbon building measures, Proc. Eng. 118 (2015) 606–613, https:// doi.org/10.1016/j.proeng.2015.08.490. [29] M. Gul, E. Celik, A.T. Gumus, A.F. Guneri, A fuzzy logic based PROMETHEE method for material selection problems, Beni-Suef Univ. J. Basic Appl. Sci. 7 (2018) 68–79, https://doi.org/10.1016/j.bjbas.2017.07.002. [30] S.O. Ajayi, L.O. Oyedele, B. Ceranic, M. Gallanagh, K.O. Kadiri, Life cycle environmental performance of material specification: a BIM-enhanced comparative assessment, Int. J. Sustain. Build. Technol. Urban Dev. 6 (2015) 14–24, https://doi. org/10.1080/2093761X.2015.1006708. [31] J. Crippa, L.C. Boeing, A.P.A. Caparelli, M. do R. de M.M. da Costa, S. Scheer, A.M. F. Araujo, D. Bem, A BIM–LCA integration technique to embodied carbon estimation applied on wall systems in Brazil, in: Built Environment Project and Asset Management, 2018, https://doi.org/10.1108/BEPAM-10-2017-0093. [32] M. Najjar, K. Figueiredo, M. Palumbo, A. Haddad, Integration of BIM and LCA: evaluating the environmental impacts of building materials at an early stage of designing a typical office building, J. Build. Eng. 14 (2017) 115–126, https://doi. org/10.1016/j.jobe.2017.10.005. [33] J. Schultz, K. Ku, M. Gindlesparger, J. Doerfler, A benchmark study of BIM-based whole-building life-cycle assessment tools and processes, Int. J. Sustain. Build. Technol. Urban Dev. 7 (2016) 219–229, https://doi.org/10.1080/ 2093761X.2017.1302839. [34] S. Azhar, W.A. Carlton, D. Olsen, I. Ahmad, Building information modeling for sustainable design and LEED ® rating analysis, Autom. Constr. 20 (2011) 217–224, https://doi.org/10.1016/j.autcon.2010.09.019. [35] V.J.L. Gan, M. Deng, K.T. Tse, C.M. Chan, I.M.C. Lo, J.C.P. Cheng, Holistic BIM framework for sustainable low carbon design of high-rise buildings, J. Clean. Prod. 195 (2018) 1091–1104, https://doi.org/10.1016/j.jclepro.2018.05.272. [36] V.J.L. Gan, J.C.P. Cheng, I.M.C. Lo, C.M. Chan, Developing a CO2-e accounting method for quantification and analysis of embodied carbon in high-rise buildings, J. Clean. Prod. 141 (2017) 825–836, https://doi.org/10.1016/j. jclepro.2016.09.126. [37] P. Inyim, J. Rivera, Y. Zhu, Integration of building information modeling and economic and environmental impact analysis to support sustainable building design, J. Manag. Eng. 31 (2014), https://doi.org/10.1061/(ASCE)ME.1943- 5479.0000308. [38] X. Yang, M. Hu, J. Wu, B. Zhao, Building-information-modeling enabled life cycle assessment, a case study on carbon footprint accounting for a residential building in China, J. Clean. Prod. 183 (2018) 729–743, https://doi.org/10.1016/j. jclepro.2018.02.070. [39] R. Frischknecht, M. Balouktsi, T. Lützkendorf, A. Aumann, H. Birgisdottir, E. G. Ruse, A. Hollberg, M. Kuittinen, M. Lavagna, A. Lupiˇ sek, A. Passer, B. Peuportier, L. Ramseier, M. R¨ ock, D. Trigaux, D. Vancso, Environmental benchmarks for buildings: needs, challenges and solutions—71st LCA forum, Swiss Federal Institute of technology, Zürich, 18 June 2019, Int. J. Life Cycle Assess. 24 (2019) 2272–2280, https://doi.org/10.1007/s11367-019-01690-y. [40] D. Trigaux, K. Allacker, W. Debacker, Environmental benchmarks for buildings: a critical literature review, Int. J. Life Cycle Assess. (2020), https://doi.org/ 10.1007/s11367-020-01840-7. [41] A. Alghamdi, G. Hu, H. Haider, K. Hewage, R. Sadiq, Benchmarking of water, energy, and carbon flows in academic buildings: a fuzzy clustering approach, Sustainability. 12 (2020) 1–25, https://doi.org/10.3390/su12114422. [42] C. Ji, T. Hong, J. Jeong, J. Kim, M. Lee, K. Jeong, Establishing environmental benchmarks to determine the environmental performance of elementary school buildings using LCA, Energy Build. 127 (2016) 818–829, https://doi.org/10.1016/ j.enbuild.2016.06.042. [43] R. Moschetti, L. Mazzarella, N. Nord, An overall methodology to define reference values for building sustainability parameters, Energy Build. 88 (2015) 413–427, https://doi.org/10.1016/j.enbuild.2014.11.071. [44] Ministry of Presidency, Orden ECO/805/2003, de 27 de marzo, sobre normas de valoraci´ on de bienes inmuebles y de determinados derechos para ciertas finalidades financieras / Economic evaluation of real state and rights for certain financial ends, 2003, p. 48. https://www.boe.es/buscar/pdf/2003/BOE-A-2003-7 253-consolidado.pdf (accessed August 10, 2021). [45] P. Mercader-Moyano, Cuantificaci´ on de los recursos consumidos y emisiones de CO2 producidas en las construcciones de Andalucía y sus implicaciones en el protocolo de Kioto (Quantification of the resources consumed and CO2 emissions produced in the constructions of Andalusia and its implications in the Kyoto protocol), Universidad de Sevilla, 2010. https://idus.us.es/bitstream/handle /11441/15660/Q_Tesis_MPMM-1.pdf?sequence=1 (accessed August 10, 2021). [46] M.P. del Mercader, M. Marrero, J. Solís, M.V. Montes, A. Ramírez, Cuantificaci´ on de los recursos materiales consumidos en la ejecuci´ on de la cimentaci´ on (Quantification of the material resources consumed in the execution of the foundation), Informes de La Construcci´ on (Constr. Rep.) 62 (2010) 125–132, https://doi.org/10.3989/ic.09.000. [47] P. Gonz´ alez-Vallejo, J. Solís-Guzm´ an, R. Ll´ acer, M. Marrero, La construcci´ on de edificios residenciales en Espa˜ na en el período 2007-2010 y su impacto según el indicador Huella Ecol´ ogica (The construction of residential buildings in Spain in the 2007-2010 period and its impact according to the Ecological Footprint indicator), Inf. Constr. 67 (2015), e111, https://doi.org/10.3989/ic.14.017. [48] KieranTimberlake Innovations, Tally Website. https://www.choosetally.com/, 2020. [49] A. Martínez-Rocamora, J. Solís-Guzm´ an, M. Marrero, LCA databases focused on construction materials: a review, Renew. Sust. Energ. Rev. 58 (2016) 565–573, https://doi.org/10.1016/j.rser.2015.12.243. [50] B.P. Weidema, M. Thrane, P. Christensen, J. Schmidt, S. Lokke, Carbon footprint: a catalyst for life cycle assessment? J. Ind. Ecol. 12 (2008) 3–6, https://doi.org/ 10.1111/j.1530-9290.2008.00005.x. [51] J. Cagiao, B. G´ omez, J.L. Dom´ enech, S. Guti´ errez Mainar, H. Guti´ errez Lanza, Calculation of the corporate carbon footprint of the cement industry by the application of MC3 methodology, Ecol. Indic. 11 (2011) 1526–1540, https://doi. org/10.1016/j.ecolind.2011.02.013. [52] UNE-EN 15978, Sustainability of construction works. Assessment of environmental performance of buildings, in: Calculation Method, 2012. http://www.aenor.es/aen or/normas/normas/fichanorma.asp?tipo=N&codigo=N0049397 (accessed August 10, 2021). [53] M.P.W. Spain, Cat´ alogo de Elementos Constructivos (Construction Elements Catalog) (CEC). https://www.codigotecnico.org/images/stories/pdf/aplicacion es/nCatalog_infoEConstr/CAT-EC-v06.3_marzo_10.pdf, 2010. [54] P. Gonz´ alez-Vallejo, Evaluaci´ on econ´ omica y ambiental de la construcci´ on de edificios residenciales (Economic and environmental evaluation of the construction of residential buildings), in: Aplicaci´ on a Espa˜ na y Chile (Application to Spain and Chile), 2017. https://idus.us.es/bitstream/handle/11441/64005/GonzalezValle joP.pdf?sequence=1 (accessed August 10, 2021). [55] A. Freire-Guerrero, M.D. Alba-Rodríguez, M. Marrero, A budget for the ecological footprint of buildings is possible: a case study using the dwelling construction cost A. Martínez-Rocamora et al. Automation in Construction 132 (2021) 103980 17 database of Andalusia, Sustain. Cities Soc. 51 (2019) 101737, https://doi.org/ 10.1016/j.scs.2019.101737. [56] M.K. Dixit, Life cycle embodied energy analysis of residential buildings: a review of literature to investigate embodied energy parameters, Renew. Sust. Energ. Rev. 79 (2017) 390–413, https://doi.org/10.1016/j.rser.2017.05.051. [57] E. Lewinson, Explaining Feature Importance by Example of a Random Forest, Towards Data Science. https://towardsdatascience.com/explaining-feature-impo rtance-by-example-of-a-random-forest-d9166011959e, 2019. [58] P. Pło´ nski, The 3 Ways to Compute Feature Importance in the Random Forest, Towards Data Science. https://towardsdatascience.com/the-3-ways-to-compute- feature-importance-in-the-random-forest-96c86b49e6d4, 2020. [59] M. Cerliani, Feature Importance with Neural Network, Towards Data Science. htt ps://towardsdatascience.com/feature-importance-with-neural-network-346eb6 205743, 2019 (accessed March 7, 2021). [60] D. Bienvenido-Huertas, C. Rubio-Bellido, J.L. P´ erez-Ord´ o˜ nez, M.J. Oliveira, Automation and optimization of in-situ assessment of wall thermal transmittance using a Random Forest algorithm, Build. Environ. 168 (2020), https://doi.org/ 10.1016/j.buildenv.2019.106479. [61] T. Parhizkar, E. Ra, A. Parhizkar, Evaluation and improvement of energy consumption prediction models using principal component analysis based feature reduction, J. Clean. Prod. 279 (2021) 123866, https://doi.org/10.1016/j. jclepro.2020.123866. [62] S. Walker, W. Khan, K. Katic, W. Maassen, W. Zeiler, Energy & Buildings Accuracy of different machine learning algorithms and added-value of predicting aggregated-level energy performance of commercial buildings, Energy Build. 209 (2020) 109705, https://doi.org/10.1016/j.enbuild.2019.109705. [63] A. G´ eron, Hands-On Machine Learning with Scikit-Learn, Keras, 2nd ed., O’Reilly Media, Sebastopol, Canada, 2019. ISBN 978-1-492-03264-9. [64] N. Bhatia, What Is out of Bag (OOB) Score in Random Forest? Towards Data Science. https://towardsdatascience.com/what-is-out-of-bag-oob-score-in-rando m-forest-a7fa23d710, 2019. [65] North American Insulation Manufacturers Association, Environmental Product Declaration (EPD) - Mineral Wool Board, 2018, p. 18. https://www.jm.com/conten t/dam/jm/global/en/building-insulation/Files/BIToolbox/Mineral-Wool-En vironmental-Product-Declaration.pdf (accessed September 14, 2021). [66] Marcel G´ omez Consultoría Ambiental, Declaraci´ on Ambiental de Producto (DAPEPD) - Plancha de aislamiento t´ ermico de espuma de poliestireno extruido (XPS) Danopren (Environmental Product Declaration (DAP-EPD) - Danopren extruded polystyrene foam (XPS) thermal insulation sheet), 2014, p. 23. http://materiales. gbce.es/wp-content/uploads/2017/07/DAP-aislamiento-t´ ermico-XPS-DANOPREN .pdf (accessed September 14, 2021). [67] PU Europe, Environmental Product Declaration (EPD) for PU (PUR/PIR) Thermal Insulation Boards and Energy Saving Potential, 2014, p. 7. http://highperforman ceinsulation.eu/wp-content/uploads/2016/08/Factsheet_13-1_Environmental_p roduct_declaration__EPD__for_PU__PUR-PIR__thermal_insulation_boards_and_ene rgy_saving_potential__updated_12-12-14_.pdf (accessed September 14, 2021). [68] J. Solís-Guzm´ an, C. Rivero-Camacho, D. Alba-Rodríguez, A. Martínez-Rocamora, Carbon footprint estimation tool for residential buildings for non-specialized users: OERCO2 project, Sustainability. 10 (2018) 1359, https://doi.org/10.3390/ su10051359. [69] C. De Wolf, Material Quantities in Building Structures and their Environmental Impact, Massachusetts Institute of Technology, June 2014. https://dspace.mit.edu /bitstream/handle/1721.1/91298/893482053-MIT.pdf?sequence=2&isAllowed =y (accessed September 14, 2021). [70] P. Chastas, T. Theodosiou, K.J. Kontoleon, D. Bikas, Normalising and assessing carbon emissions in the building sector: a review on the embodied CO 2 emissions of residential buildings, Build. Environ. 130 (2018) 212–226, https://doi.org/ 10.1016/j.buildenv.2017.12.032. [71] P. Gonz´ alez-Vallejo, R. Muntean, J. Solís-Guzm´ an, M. Marrero, Carbon footprint of dwelling construction in Romania and Spain. A comparative analysis with the OERCO2 tool, Sustainability (Switzerland) 12 (2020), https://doi.org/10.3390/ SU12176745. [72] Andalusia Government, Andalusia Construction Cost Database (ACCD). http:// www.juntadeandalucia.es/organismos/fomentoyvivienda/areas/vivienda-rehabili tacion/planes-instrumentos/paginas/bcca-sept-2017.html, 2017 (accessed March 21, 2018). [73] M. Marrero, M. Wojtasiewicz, A. Martínez-Rocamora, J. Solís-Guzm´ an, M.D. Alba- Rodríguez, BIM-LCA integration for the environmental impact assessment of the urbanization process, Sustainability (Switzerland). 12 (2020) 1–24, https://doi. org/10.3390/su12104196. [74] K. Goulouti, P. Padey, A. Galimshina, G. Habert, S. Lasvaux, Uncertainty of building elements’ service lives in building LCA & LCC: What matters? Build. Environ. 183 (2020) https://doi.org/10.1016/j.buildenv.2020.106904. [75] O. Larsson Ivanov, D. Honfi, F. Santandrea, H. Stripple, Consideration of uncertainties in LCA for infrastructure using probabilistic methods, Struct. Infrastruct. Eng. 15 (2019) 711–724, https://doi.org/10.1080/ 15732479.2019.1572200. [76] X. Zhang, R. Zheng, F. Wang, Uncertainty in the life cycle assessment of building emissions: a comparative case study of stochastic approaches, Build. Environ. 147 (2019) 121–131, https://doi.org/10.1016/j.buildenv.2018.10.016. [77] X. Zhang, K. Liu, Z. Zhang, Life cycle carbon emissions of two residential buildings in China: comparison and uncertainty analysis of different assessment methods, J. Clean. Prod. 266 (2020) 122037, https://doi.org/10.1016/j. jclepro.2020.122037. [78] M.F.D. Morales, N. Reguly, A.P. Kirchheim, A. Passuello, Uncertainties related to the replacement stage in LCA of buildings: a case study of a structural masonry clay hollow brick wall, J. Clean. Prod. 251 (2020) 119649, https://doi.org/10.1016/j. jclepro.2019.119649. A. Martínez-Rocamora et al.