Automatic BIM detailing using deep features of 3D views
Sou-Han Chen 1 and Fan Xue 2
This is the peer-reviewed post-print version of the paper:
Chen, S.-H., & Xue, F. (2023). Automatic BIM detailing using deep features of 3D
views. Automation in Construction, 148, 104780. Doi:
10.1016/j.autcon.2023.104780
The final version of this paper is available at: https://doi.org/10.1016/j.autcon.2023.104780.
The use of this file must follow the Creative Commons Attribution Non-Commercial No
Derivatives License, as required by Elsevier’s policy.

5

10

15

20

Abstract
Building information modeling (BIM) detailing, the process of adding the level of graphical
and non-graphical details, is required in many BIM stages and applications; however, manual
BIM detailing is a resource-intensive and costly process. This study proposes an automatic
BIM detailing method based on deep features (DFs) of BIM 3D views in three steps. First, a
BIM’s 3D view and semantics were extracted automatically. Then, machine learning (ML)
algorithms learned the DFs to predict the target BIM’s invisible details. Finally, the details
were automatically added to BIM by a Dynamo program. A case study of motion-bearing
component detailing for 86 doors through three DFs and five ML algorithms revealed that
DFs improved the automatic detailing results comprehensively (29 out of 32 scenarios) and
significantly. This paper’s contribution includes an effective, novel approach for automatic
BIM detailing as well as quantified experimental evidence about the effectiveness of DFs for
BIM applications.
Highlights
• An automatic BIM detailing method using machine learning (ML) and deep features (DFs).
• A case study of motion-bearing component detailing for 86 door types in NBS BIM Library.
• Experiments conducted for 7 detailing tasks using 5 ML algorithms and 3 DFs.
• Comprehensiveness: 29 out of 32 ML-task combinatorial scenarios improved by DFs.
• Significance: averagely 47.7% error reduction for classification and 18.8% for regression.
Keywords
Building information modeling; BIM detailing; deep features; motion-bearing components;
NBS BIM library
Sou-Han CHEN, Research Assistant
Department of Real Estate and Construction, The University of Hong Kong, Pokfulam, Hong Kong SAR, China
Email: souhan@hku.hk, ORCiD : https://orcid.org/0000-0003-0511-7518
1

Fan XUE, Assistant Professor, Ph.D.
Department of Real Estate and Construction, The University of Hong Kong, Pokfulam, Hong Kong SAR, China
Email: xuef@hku.hk, ORCiD : https://orcid.org/0000-0003-2217-3693
2

*: Corresponding author, Tel: +852 3917 4174, Fax: +852 2559 9457, Email: xuef@hku.hk

25

30

35

40

45

50

55

60

1 Introduction
The architecture, engineering, construction, and operation (AECO) sector is pursuing a digital
transformation to evolve the traditional workflow (NIC 2017; CIC 2021). Building
information modeling (BIM) is ‘a comprehensive digital representation of physical and
functional characteristics of a facility’ (NIBS 2015). In addition to the geometric 3D model,
BIM can store various types of non-geometric data, relationships, and documentation. The
benefits and impacts of BIM adoption have been agreed upon and discussed by researchers
worldwide (Barlish & Sullivan 2012; Ghaffarianhoseini et al. 2017). BIM has had a positive
impact on many construction applications, such as workflow standardization, information
integration, project management, and data analytics. Therefore, the AECO industry has
advocated and promoted the importance of mandating BIM around the world (Chan et al.
2019) and is gearing up with online and offline BIM libraries (Kim et al. 2015).
BIM detailing refers to the process of increasing the level of detail (LOD) for construction
projects. Unlike the level of development (LoD), referring to the development of reliable and
concrete BIM at different lifecycles of the project, LOD refers to the accurate details of BIM.
In general, the development of the LOD of BIM will increase progressively as the project
moves on. The design stage will primarily have BIM with lower LOD, while during the
construction period, BIM will be further enhanced and detailed by engineers from different
disciplines. The process of BIM detailing is labor-intensive and time-consuming, leading to a
need for an automation process. Currently, the performance and applications of BIM heavily
depend on the quality, semantics, and LOD of BIM. Meanwhile, research also emphasized
the importance of information depth and pointed out the consequences of information loss
(Borrmann et al. 2018).
BIM detailing is vital to the BIM lifecycle. The insufficient LOD of BIM can lead to the
failure of performing time dynamic simulation for digital twin applications. A digital twin is
a virtual representation of a physical object or system across its lifecycle, using real-time data
to enable understanding, learning, and reasoning (NIC 2017). The International Council for
Building also suggests that the digital twin is a prominent topic in the AECO sector as it
provides a cyber-physical integration of the dynamic building status, including the
construction and operation phase (Seaton et al. 2022). For example, a door without detailed
hinges falls rather than opens after a simulated push, as shown in Figure 1. In contrast, the
constraint from hinges leads to a correct motion behavior in a digital twin. Nevertheless, due
to the low added value of manually created high LOD of BIM objects, many BIM objects
might not have adequate components of information needed in the digital twin applications.
One distinct example would be the motion-bearing components. Without the motion-bearing
components, the digital twin might not be able to accurately present and visualize the
collected real-time data. Therefore, an efficient method for automatic BIM detailing should
be developed.

65

Figure 1. A failed door simulation resulting in falling without hinge details in the BIM

70

75

Machine learning (ML), including deep learning, has been demonstrated to be a set of
effective approaches to facilitate and automate BIM detailing and enrich construction
information. ML and deep learning can be used to automate and optimize BIM applications
(Zabin et al. 2022) and predict missing construction data (Yang et al. 2021). Recent studies
(Dargan et al. 2020) showed that deep features (DFs) could be successfully integrated into
many applications (e.g., image recognition, semantic image segmentation, and object
detection). The results of such applications can be further applied in assisting BIM parametric
design through various tools. For example, Dynamo from Autodesk Revit provides a visual
programming environment to quickly perform a data-driven parametric BIM design process
(Kensek 2015).

90

This study presents an automatic BIM detailing approach through the exploitation of the
target BIM’s DFs. To validate the presented approach, we collected a set of training BIM
objects from the NBS National BIM Library (NBS 2022). Based on the learning results of
DFs in the training of BIM objects, the missing BIM components can automatically be
predicted as three classification tasks and four regression tasks; and created via a Dynamo
program. The contribution of this paper is two-fold. First, from the perspective of knowledge,
the outcomes of this paper confirmed that DFs could considerably facilitate the ML-based
BIM detailing; for example, all the classification tasks were improved by 30.4~58.7% on
average, while four regression tasks were improved by 9.5~33.6%. For BIM practitioners,
automatic detailing based on the BIM sketch design can escalate productivity and save costs
in terms of human resources and machine hours.

95

Following this introduction, Section 2 provides a literature review of BIM, detailing deep
learning applications for BIM. Section 3 presents a proposed three-step BIM detailing
approach, that is, feature extraction, component parameters prediction, and parametric design.
Section 4 shows a case study of detailing door hinges with a training set of 86 BIM objects,

80

85

together with evaluations and analyses of various deep features and ML methods. Section 5
discusses the findings and potential impacts, along with the limitations. Section 6 concludes
this study and points out future works.
100

105

110

115

120

125

130

2 Literature review
2.1 BIM detailing
BIM has been increasingly adopted in the AECO sector. On the one hand, BIM provides a
systematic information hub for various project practitioners to collaborate at the same pace.
On the other hand, BIM is considered a ‘blind and deaf’ model as the information stored
inside cannot be easily synchronized with actual built assets as the project progresses (Chen
et al. 2015). To alleviate such static, outdated, and lacking detail issues in BIM, the concepts
of as-built BIM and digital twin building emerged (Wu et al. 2021). The as-built BIM
provides up-to-date information on the current built assets (Xue 2022), while the digital twin
provides a systematic and dynamic virtual environment for retrieving, simulating, and
manipulating different levels of information. However, there are two main gaps between BIM
and digital twin: the LOD of the model and the capability of handling the dynamic data
collected from reality (Opoku et al. 2021).
To bridge the gaps between a ‘blind and deaf’ BIM and digital twin, academic research and
industrial applications in the literature have reported three aspects of BIM detailing, as
contrasted in Table 1. An example group of research is BIM integration with other
technologies, such as light detection and ranging (LiDAR) (Xue et al. 2019), Internet of
Things (IoT) (Tang et al. 2019), geographic information system (GIS) (Wang et al. 2019),
and machine learning (ML) (Zabin et al. 2022); BIM data exchange and interoperability (Lou
et al. 2020), and optimization (Lou et al. 2020); BIM interactive application through virtual
reality (VR) (Zhang et al. 2021) and gaming technology (Potseluyko et al. 2022). On the
other way, as the guidelines and standards of BIM have been developed by both international
and local institutions (CIC 2020; ISO 2018), the application of BIM in the industry is
booming and based mostly on the stage of the project: the design stage focuses on the
simulation and planning of the models; the construction stage focuses on BIM detailing, 4D
BIM, and project monitoring; and the operation stage focuses on the documentation and
maintenance of the project. As construction projects by nature are dynamic and contain lots
of variables and uncertainties, the industry is seeking a more dynamic BIM to reveal an
accurate and instant project status, such as motion and load analysis during the design and
construction stage.
Table 1. BIM detailing research direction
BIM detailing aspects
Semantic enrichment
Semantic constitution
Semantic implementation

Research direction
BIM integration
BIM interoperability & optimization
BIM simulation & visualization

Technologies adoption
LiDAR, IoT, GIS, ML
BIM
VR, gaming technology

135

140

145

150

155

160

165

170

The detail of the model is critical to the successful implementation of BIM in managing
dynamic projects. A higher LOD of BIM enables different applications to perform more
dynamically and store more semantic information in the models, which can further affect and
assist stakeholders in the decision-making process (Boje et al. 2020) as well as provide
engineers a comprehensive project status (Lu et al. 2020). However, as BIM detailing is timeconsuming and requires interdisciplinary engineers to create detailed components and
collaborate throughout the project, an appropriate LOD of BIM would be preferred and
suggested based on different stages of the project (Fai & Rafeiro 2014). Similarly,
researchers suggested different BIM applications should imply different LOD to achieve cost
savings (Hong et al. 2019). Therefore, some components and items are often neglected during
BIM detailing, as they will not affect the construction process or technical analysis or conflict
with any stakeholder’s interest.
2.2 Machine learning and deep learning applications for BIM
Machine learning (ML) refers to the scientific study of algorithms and models that enable
computer systems performing certain tasks without predefined and explicitly programmed
processes (Han et al. 2011). These algorithms and models can be categorized into different
groups based on how they are being trained and their applications. For applications in the
AECO sector, there are three categories, instance-based ML, rule/tree-based ML, and
function-based ML (Hall et al. 2009) that are commonly used for data prediction, image
processing, simulation, and modeling. In all, the conventional ML techniques require
structured and well-organized input data.
Deep learning (DL) techniques, in contrast, allow models to learn representations of raw data
with multiple levels of abstraction (LeCun et al. 2015). The capabilities of deep learning to
analyze data, provide insights, optimize workflows, and make predictions have become a new
method for the AECO sector to process BIM data across the BIM lifecycle. For the design
stage, deep learning has been proven to be an effective method to perform generative design
(Bianconi et al. 2019; Chen et al. 2022), clash prediction along with resolution (Hu & CastroLacouture 2019), and compliance checking (Song et al. 2018) by manipulating BIM data. For
the construction stage, deep learning can provide cost estimation (Banihashemi et al. 2022)
and optimize project schedules (Torres-Calderon et al. 2019) by fully utilizing the
information stored in BIM. Lastly, during the operation stage, deep learning can be used to
classify the building type (Lomio et al. 2018), enrich semantic information in the as-built
BIM (Xue et al. 2021; Xue 2022), automatically conduct the property valuation (Su et al.
2021), and assist the defect surveying (Valero et al. 2018). As deep learning progressively
develops, more potential applications for BIM will be explored.

175

180

185

190

195

200

Since current technologies of deep learning cannot directly use native BIM-format data as the
training data and output data, an appropriate data exchange method should be adopted for
BIM. Currently, many BIM plug-ins have been developed and can fill this gap, such as
Grasshopper for Rhinoceros, PARAM-O for ArchiCAD, and Dynamo for Revit. These plugins enable users to use the external data for parametric design and export the needed data for
analyses (Natanian et al. 2019). Take Dynamo as an example; it provides a graphical
interface for users to perform parametric-related functions of BIM objects (Kensek 2015).
Research also shows that Dynamo can effectively facilitate complex model generation (Yang
et al. 2020) and extract data for operational usage (Sadeghi et al. 2019; Thabet et al. 2022).
Another illustrated the adoption of Dynamo to facilitate the process of 3D concrete printing
(Weng et al. 2021).
In summary, as the adoption of BIM increases in the AECO sector, various BIM applications
have been explored and adopted rapidly, including the digital twin of the project. However,
the current industrial practice requires great resources to achieve a higher LOD of BIM to
perform dynamic and accurate project status. Meanwhile, the rapid development and
application of deep learning technologies in BIM have proven to be effective in facilitating
BIM-related works. As a result, inspired by previous research, this project will utilize deep
features to fill in the missing motion components in BIM objects to increase the interactive
and dynamic performance of BIM applications.
3 The proposed automatic BIM detailing method
3.1 Overview
The proposed method aims to automatically create detailed BIM components with the deep
features of 3D views. Figure 2 shows the developed three-step method for the automated
generation of BIM details. The three steps are (1) characteristic 3D views of BIM, (2) deep
learning of BIM details, and (3) parametric BIM detailing. Each step has a set of inputs,
which in part depend on the prior step.

Step
(Subsect.)

Inputs

Process

Step 1
(§ 3.2)

BIM objects with
details

3D view and semantics
extraction in Dynamo

3D view images
and BIM semantics

Well-known imagery
deep features

Comparative study of the
deep features

Evaluation and
selection

Training with selected deep
features

Trained BIM
detailing models

3D view extraction in
Dynamo; and prediction

Predicted BIM
details

Automatic BIM detailing in
Dynamo

BIM objects
enriched with details

Step 2
(§ 3.3)

Step 3
(§ 3.4)

BIM objects
(without details)

Outputs

So�ware

Figure 2. Method of the proposed three-step automatic BIM detailing using deep features
205

210

215

3.2 BIM 3D views and semantics extraction
The first step extracts the semantics and 3D views of an input BIM for the subsequent deep
learning processes. BIM can be considered a well-structured information hub (Heaton et al.
2019). However, the BIM software ecosystem has unsatisfactory interoperability with
existing ML and deep learning algorithms in the computer science field. Therefore, the
semantics and 3D view imagery of an input BIM must be converted from the BIM object to
machine learning-friendly formats. Given a BIM object with expected details in BIM
software, users can use the built-in functions or plug-ins or self-programming to generate the
data for machine learning use. Using Dynamo, a Revit plug-in for extracting data from the
BIM family, as an example, the general steps are shown in figure 3. First, the Dynamo scripts
enable Revit to load family documents in batches. The 3D views of the BIM objects can be
generated by image export Revit, or users can predefine the angle and export images through
Dynamo or self-programming in C#. Meanwhile, the semantic data can be filtered and
retrieved through Dynamo or using Revit API. Finally, 3D views in PNG format and
parameters in CSV format are extracted, stored, and ready for the next step.

220

Figure 3. Data conversion from BIM to ML

225

3.3 DF-based BIM details prediction
This step aims to train the ML algorithms for the capability of predicting the parameter
values of BIM objects. To achieve this target, the adoption of DFs was proposed during the
ML training process. Three well-known DFs, VGG19, InceptionV3, and SqueezeNet, are
investigated in this study. DFs of the images can represent the unstructured image in a tabular

230

format. VGG19 is a typical 19-layer deep convolutional neural network (CNN) proposed in
2014 and widely used in the medical sector for disease diagnosis (Alhindi et al. 2018).
InceptionV3 (InceptV3) is a 48-layer deep CNN proposed in 2015 that uses a splitting
method to effectively extract the DFs (Dong et al. 2020). SqueezeNet is a 48-layer deep CNN
proposed in 2016 that uses fewer parameters for the feature extraction with competitive
accuracy and smaller model sizes (Lee et al. 2019).

235

240

245

250

As shown in Figure 2, a comparative study is conducted to select the most suitable model.
The DF-based learning process of the model includes two types of features, 3D views’ DFs
and BIM semantic features. The prediction targets of BIM details can be broadly classified
into two types: discrete values for classification and continuous values for regression.
Examples of classification are types of family and materials of parts, while regression is the
locational values of the components. This study adopts five representative ML algorithms
from three ML categories, as listed in Table 2. K-nearest neighbors (KNN) algorithm is
selected as a representative in the instance-based ML category. From the rule/tree-based
category, decision tree (Tree) is a typical and well-known ML algorithm, while the eXtreme
Gradient Boosting (XGBoost) is selected as a more advanced algorithm that exploits the
second order gradient statistics of the loss function. From the function-based category, linear
regression (LR) and support vector machine (SVM) are selected for their linear relationship
and kernel function models.
Table 2. Selected ML algorithms for BIM details classification and regression
ML category
Instance-based
Rule/tree-based
Function-based

255

ML algorithm
KNN
Tree
XGBoost
LR
SVM

Classification
Y
Y
Y
N
Y

Regression
Y
Y
Y
Y
Y

Reference
(Zhang & Zhou 2007)
(Han et al. 2011)
(Chen & Guestrin 2016)
(Han et al. 2011)
(Hearst et al. 1998)

Both classification and regression training processes are based on 5-fold cross-validation. To
evaluate the performance of the trained models, F1 will be used for the comparison of
classification algorithms, while RMSE will be used for the comparison of regression
algorithms. Table 3 lists the error metrics and selection in the two processes.
Table 3. Error metrics for BIM details classification and regression
Task
Classification

Name
Precision

Regression

Recall
F1
RMSE

Definition
the proportion of true positives among instances classified as
positive
the proportion of true negatives among all negative instances
the weighted harmonic mean of precision and recall
square root of the arithmetic mean of the squares of a set of
numbers

Selected?
N
N
Y
Y

260

265

270

275

280

A series of comparisons and evolutions of trained models is conducted based on the error
metrics. First, whether or not the model is improved can be gleaned by comparing the best F1
value for classification or best root mean square error (RMSE) value for regression between
the control group (i.e., without DF) and the experimental group (i.e., with DF). If the model is
improved, the CNN DF and ML algorithm are the best settings for the corresponding tasks. In
contrast, if the model is not improved, the adoption of CNN DF might not be suitable for the
tasks, and the setting of the ML algorithm becomes the best setting. Once the best training
settings are selected, users can input all the existing features – both 3D views and BIM
semantics – to train the model. The output of this step is a trained BIM detailing model,
which can predict the target BIM detail using 3D views and BIM semantics.
3.4 Parametric BIM detailing
This step aims to enrich the BIM objects by adopting the trained model and the predicted
information of the target details. Therefore, the input of this step is a BIM object without
target details. Firstly, the 3D views and semantics of the BIM object without details need to
be extracted as described in Section 3.2. Then users can apply the trained model in Section
3.3 to predict the discrete or continuous values of the target BIM details.
The predicted values are used to create the BIM objects parametrically, which can be done by
programming and plug-ins as shown in Figure 4. For example, Dynamo provides a node
where users can create a cylinder by inputting the radius and height; Grasshopper provides a
node where users can create a cone by inputting the base plane, the radius at the cone base,
and the cone height. Such parametric design can achieve the same goals with less time and
effort than manual drawing, especially with many duplicate tasks. Eventually, the output of
this step is the enriched BIM objects with details.

285

Figure 4. General framework of Dynamo program

290

4 A study of NBS doors
4.1 Case study
A case study of invisible hinges for swing door families in the UK National Building
Specification (NBS) BIM Library was conducted to advance the proposed method. Doors are
one of the most common and movable object types in BIM projects, and hinges are the
critical motion-bearing details. In the NBS BIM Library, 48 of the 94 door families,

295

consisting of 86 door types according to the parametric settings, conveyed critical details of
hinges. The proposed method was tested for predicting and detailing the motion-bearing
components, i.e., invisible hinges. The training BIM object set thus was the collection of all
86 door types with hinges. It was a challenging task to detail the hinges correctly for various
door types automatically; it was even more challenging in this case since hinges are invisible
from the training external 3D views.

300

305

Table 4 lists the target details, including quantity, depth, radius, and different positions, to
predict hinges. ML algorithms were trained using native BIM semantics and 3D view’s deep
features. Table 5 lists the three general parameters, NominalHeight, NominalLength, and
NominalWidth, of doors defined by NBS BIM Object Standard v2.1 for training the ML
algorithms.
Table 4. List of targets to predict for hinge detailing
Task Type
Classification

Task name
Quantity
Depth
Radius
Regression
Position (Top)
Position (Bottom)
Position (Extra1)
Position (Extra2)*
*Only applicable to the 4-hinge door.
310

Table 5. Selected BIM parameters
Property name
NominalHeight
NominalLength
NominalWidth

315

Definition
The quantity of hinge. (either 3-hinge or 4-hinge)
The height of the hinge (cylinder).
The radius of the hinge (cylinder).
The ratio of the highest hinge Z-position to the height of the door.
The ratio of the lowest hinge Z-position to the height of the door.
The ratio of the 2nd-lowest hinge Z-position to the height of the door.
The ratio of the 3rd-lowest hinge Z-position to the height of the door.

Property requirements
A numerical value of the nominal height (typically the vertical characteristic dimension of
the product) in millimeters.
A numerical value of the nominal length (typically the primary or larger of the two
perpendicular horizontal dimensions of the product) in millimeters.
A numerical value of the nominal width (typically the secondary or smaller of the two
perpendicular horizontal dimensions of the product) in millimeters.

4.2 Experimental settings
The experiments were conducted on a MacBook Pro M1 with a CPU (3.2 GHz, 8-core), 16
GB memory, and MacOS (ver. 12.4). The test software included Autodesk Revit (ver. 2022),
Dynamo (ver. 2.12), and Orange data mining (ver. 3.31), as listed in table 6.
Table 6. Experimental environment setting
Software
Revit
Dynamo

Version
2022
2.12

Functionality
Loading Revit family documents (rfa.) file and supporting dynamo
BIM semantics and 3D view image extraction; parametric design for hinge
creation

Orange

320

325

3.31

Deep feature extraction, machine learning model training, evaluation, and
prediction

Figure 5 shows the Dynamo scripts for extracting the door types’ 3D views and semantics,
respectively. Figure 5(a) shows the script for exporting 3D views. The view angle was set as
the top right corner of the exterior side of the door. The image resolution was set to be 1024 x
1024 pixels at 600 DPI. For standardized qualities of the view images, the ‘detail level’ on
the Revit view control bar was selected as ‘fine,’ the view scale was set as 1:10, and the
visual style was set as ‘hidden line.’ Also, all the annotation categories and the walls model
categories were disabled inside the visibility/graphic setting. The final output of the image is
the 3D view of the exterior side in PNG format. As shown in Figure 5(b), the three standard
NBS parameters, NominalHeight, NominalLength, and NominalWidth, were exported into an
index file in the CSV format. Meanwhile, hinge-related BIM parameters (i.e., the quantity,
depth, radius, and relative locations) were also extracted into the index file.

330

(a)

(b)

Export 3D Views

Load Revit Families

Retrieve parameter values

Export to CSV

Figure 5. Dynamo scripts in Step 1. (a) Exporting 3D views; (b) Parameter extraction

335

340

Figure 6 shows the workflow diagrams of the software Orange for extracting the DFs of the
3D view images and comparisons of ML settings. For classification tasks of the quantity,
radius, and depth of the hinge, Figure 6(a) shows four ML algorithms were involved, while
all five ML algorithms were tested for regression, as shown in Figure 6(b). The left dashed
box in blue color represents the extraction of the DFs from 3D view images, while the right
dashed box in orange represents the ML process.

(b)

(a)

Deep features

ML process

Deep features

ML process

Figure 6. Orange program of deep features extraction and ML for predicting hinge details. (a)
Classification; (b) Regression
345

Once the ML model is trained, it can be used to predict the hinge-related information, as
shown in Table 3. The predicted data can be stored in a CSV file and used for the parametric
design task. Figure 7 illustrates a Dynamo script retrieving the predicted data stored in a CSV
file to perform the parametric design task for auto-enhancing the BIM object.
Door parameter extraction
Hinge generation
Importing predicted data from CSV

Create family type for hinge

350

Figure 7. Dynamo script of adding and detailing new hinge instances to the door families

355

360

4.3 Results
4.3.1 Results of Step 1
The results of Step 1, as shown in Figure 2, contained 86 view image files (.PNG format) and
an index file (.CSV format). The naming system of the PNG files is ‘Document name_family
type – 3D view – 3DExterior’, as shown in Figure 8. The resolution of each PNG file was
1024 × 1024 pixels, and the average file size was around 10KB, while the disk size of the
index file was about 20 KB. All the 3D views were shot from outside of the doors, as shown
in Figure 8; therefore, the hinges – three or four instances – were invisible in the 3D view
photos. Plus, the variety of the designs of the test door families, that is, all the three-hinge and

four-hinge doors with hinges at NBS had well covered the common door types in real
projects.

365

Figure 8. Screenshots of sample extracted 3D view images

370

375

Table 7 lists an excerpt of the decision parameters and target values extracted by the Dynamo
script automatically. In general, there were three main categories: NBS standard parameters,
discrete hinge details to classify, and continuous hinge detail values to regress. An example
of discrete hinge detail was the radius in millimeters, where the four possible values were 3.5,
6, 7.5, and 8. For the continuous hinge details to regress, the values were normalized between
0 and 1, for example, as the relative Z-position against the NominalHeight.
Table 7. Excerpt of the training data for door features and prediction results
NBS standard parameters
Nominal Nominal Nominal
Height
Length
Width
2079
96
1125
2479
96
1125
2079
96
1925
2479
96
1925
…
…
…
2110
55
1004

Classification results
Quantity Radius Depth
in mm in mm
3
3.5
100
3
3.5
100
4
3.5
100
4
3.5
100
…
…
…
4
8
100

Top

Regression
Bottom Extra1

Extra2

0.0981
0.0823
0.0981
0.0823
…
0.1123

0.8389
0.8649
0.8389
0.8649
…
0.8517

N/A
N/A
0.7186
0.7640
…
0.6242

0.4685
0.5543
0.4685
0.5543
…
0.3967

380

4.3.2 Results of Step 2
Table 8 summarizes the results of utilizing DFs in the hinge prediction process. The uniform
check marks for the three classification tasks show that DFs 100% improved the BIM details
predictions in terms of F1 scores against every baseline without DF (w/o DF). In the four
regression tasks, 17 out of 20 DF-ML combinations showed accuracy improvement
(indicated as check marks) in terms of reduced RMSE values. Thus, it was confident to
conclude that DFs can improve the prediction of the hidden BIM details.

385

Table 8. Deep feature’s improvement matrix under different ML algorithm
ML
algorithm
KNN
Tree
XGBoost
LR
SVM
Sub total

Classification prediction tasks
Quantity
Depth
Radius
✓

✓

✓

✓

✓

✓

✓
N/A
✓
4

✓
N/A
✓
4

✓
N/A
✓
4

Regression prediction tasks
Top
Bottom
Extra1
✓

✓

✓

✕

✓

✓

✕
✓
✓
3

✓
✓
✓
5

✓
✓
✓
5

Extra2
✓

Sub
total
7

✕

5

✓

4

✓

6

✓

7

4

“✓” indicates the result is improved compared to w/o DF; “N/A” indicates not applicable

390

The rightmost column in Table 8 shows that three ML algorithms (KNN, LR, and SVM)
received consistent prediction improvements in all the possible tasks. In contrast, decision
tree and XGBoost show encouraging – though not as consistent as others – results in
receiving improvements in five and six out of the seven tasks. Based on our observation and
experience during the experiments, the main reason would be the quantity (thousands of
columns) of DFs sometimes exceeded the capacity of logical conditioning/branching
combinations for rule/tree-based ML algorithm, leading to convergence at “local optima.”

395

400

405

Table 9 shows the average and the best performances of classification tasks in terms of F1
values, while the detailed result data is available in the data set in the appendix. In general,
the adoption of DFs can increase the F1 values considerably, by 15.4~58.7% on average. For
the classification tasks, ML models with DFs have considerably higher best and mean F1
scores than those w/o DF. And when comparing the mean F1 score from all ML models using
different CNNs to the mean score w/o DF, the error reduction can be calculated. In the
classification tasks, all the values of error reduction are positive, especially for the depth of
the hinge prediction task, for which the error reduction can reach 58.7%.

Table 9. Deep feature performance matrix of the classification tasks
Performance of four ML algorithms using three DFs
Best setting
*
*
#
Best F1
%Imp
Avg F1
%Imp
Stdev
Quantity
0.8951
40.0
0.8152
30.4
0.0581 InceptV3+SVM
Depth
0.9656
66.9
0.9185
58.7
0.0356 VGG19+SVM
Radius
0.9654
66.0
0.9139
15.4
0.0389 VGG19+SVM
Average
57.6
47.4
*
%Imp: Improvement in percentage calculated as error reduction in percentage, where Error = 1 − F1
#
Stdev: Standard deviation of the sixteen F1 scores from the combinations of all ML models and DFs
Tasks

410

415

w/o DF
Best F1 Avg F1
0.8252
0.7346
0.8963
0.8024
0.8983
0.8168

Table 10 shows the average and the best performances of regression tasks in terms of F1
values. For the regression tasks, the adoption of DFs reduced the RMSE values of most ML
models. However, an exception was the “Top” hinge locational prediction task, where the
best RMSE w/o DF (predicted by XGBoost) was better. The main reason should lie in the
limited capability of XGBoost in processing thousands of columns. Other models still have a
better performance, especially for the bottom hinge locational prediction task (see “Bottom”
in Table 10); the error reduction can reach 33.6% on average.
Table 10. Deep feature performance matrix of the regression tasks
w/o DF
Best
Avg
RMSE
RMSE
0.0074
0.0213
0.0093
0.0168
0.0623
0.0752
0.0374
0.0416

420

425

430

Performance of Five ML algorithms using three DFs
Best setting
Best
Avg.
%Imp#
%Imp#
Stdev^
Tasks
RMSE
RMSE
Top
0.0081
-10.2
0.0191
10.1
0.0191 InceptV3+XGBoost
Bottom
0.0079
15.7
0.0112
33.6
0.0112 VGG19+LR
Extra1
0.0388
37.7
0.0585
22.2
0.0585 InceptV3+LR
Extra2
0.0280
25.2
0.0376
9.5
0.0376 SquNet+SVM
Average
17.1
18.8
#
%Imp: Improvement in percentage calculated as error reduction in percentage, where Error = RMSE
^
Stdev: Standard deviation of the twenty RMSE scores from the combinations of all ML models and DFs

In addition to the performance, the time cost of DFs adoption was also considered and
evaluated. In this experiment, there are mainly two parts of the time cost, DFs preparation
and ML training time. For the preparation of DFs, Dynamo was used to export the 3D views
of the BIM doors. The development of the Dynamo script would take around one hour for
experienced developers or two to three hours for those not familiar with Dynamo. After
exporting the images, Orange will be used for the DFs generation and ML training. The time
performances of each DF are listed in Table 11. When comparing the effectiveness of the ML
performance enhancement and the time consumption of the training process, SqueezeNet
would be the most cost-effective CNN.

Table 11. Time performance in seconds for adopting different DFs on a laptop computer
DF
InceptV3
SqueezeNet
VGG19
435

440

DF generation (s)
2.88
2.84
2.94

Avg. classification training (s)
9.92
5.76
10.72

Avg. regression training (s)
24.36
13.30
31.18

4.3.3 Results of Step 3
Figure 9 shows the results of Step 3 for two test BIM doors for illustration. In each subfigure,
the left part is the door before automatic hinge detailing, while the right half shows the final
results of the proposed method. It can be seen that the newly-added hinges in red color were
successfully added to the original BIM door family through the presented DF-facilitated
automatic BIM detailing process.
(a)

(b)

Figure 9. The automatic BIM detailing results for hinges. (a) a single door; (b) a double door
445

450

Figure 10 compares the detailing results between different methods, including ground truth
data in black, predicting data from best ML settings with DFs (in red) and best ML settings
without DFs (in blue). In Figure 10, the DFs here did not change the quantity of the hinges,
while the differences were in the regression tasks of the Z-positions of the hinges. Since the
figures were token in the elevation view, the correct parts of the hinges were overlapped by
the ground truth. As a result, the red areas represent the errors from BIM detailing with DFs,
while the blue areas represent the errors from BIM detailing without DF. In both cases, the

considerably smaller red areas indicate that deep features led to more accurate BIM detailing
of the hinges.

455

Figure 10. Errors of hinges’ Z-positions by different ML methods. (a) a single door; (b) a
double door

460

465

4.4 A behavior simulation of the detailed doors as postprocessing
Figure 11 illustrates how the newly-detailed hinges can create more realistic scenes for BIM
simulations and visualizations. The BIM doors with newly-detailed hinges were exported
from Revit (as FBX files) to Blender through an add-on. Physics engine-based interactions
and simulations can then be created by assigning the hinges as the rigid body constraint.
When external forces are applied to the door in a BIM, hinges can prevent doors from
dropping out of the frame and can swing as expected, as shown in Figure 11. In comparison
with Figure 1, with the newly detailed motion-bearing components inside BIM, the behavior
simulations became more realistic, and the simulation results could provide the opportunity to
open the door in the virtual environment (e.g., for digital twins).

470

Figure 11. Door simulation with detailed motion-bearing components using Blender

475

480

485

490

495

5 Discussion
The method presented in this paper provides a novel workflow of the automatic BIM
detailing process, including extracting existing BIM data, adopting DFs and ML models for
predicting missing BIM components, and using parametric design to detail the missing
components automatically. The proposed method has three advantages. First, compared with
conventional manual BIM detailing, the method shortens the time of BIM detailing,
especially for complex BIM projects and hidden details, through up-to-date DFs, ML
algorithms, and parametric designs. DFs are proven successful for automatic BIM detailing
(i.e., an average of 47.4% error reduction in classification tasks and 18.8% error reduction in
regression tasks). The main reason should be the BIM design features, i.e., the door design
features like glass panels, kick plates, and handles in Figure 8, that are represented as DFs
(Han et al. 2019). These DFs can be recognized and analyzed by different ML algorithms,
and eventually improving the automatic BIM detailing process. Secondly, the general
methods are applicable to all BIM objects. Since DFs can capture the design features of the
objects, which supplement more information than typical parameters, such as Width, Height,
and Depth, to any ML algorithms, the significance of this study is not only applied to motionbearing components of the door but also other BIM components with design features in 3D
views. Lastly, the predicted parametric component details are concise and compatible with
BIM versioning such as on blockchains (Xue & Lu 2020; Zhao et al. 2023). The effects of
deep features on BIM detailing also shed light on new directions of BIM semantic enrichment
and digital twin building (Xue et al. 2021).
However, there are two main limitations in this paper, i.e., the quantity of test BIM objects
and the diversity of DFs. Although all the BIM doors annotated with nominal sizes were
downloaded from NBS BIM Library for the tests (see Table 5), the diversity of test BIM
objects was still limited. Future researchers are encouraged to test more BIM objects.

500

505

510

515

520

525

530

535

Furthermore, all three DFs in this paper were CNN DFs. Therefore, another research
direction is to examine the effectiveness of more DL/DFs, such as deep transfer learning or
transformer, for automatic BIM detailing processes.
6 Conclusion
The insufficient LOD hinders BIM applications for dynamic and interactive digital twins.
This paper presents an automatic BIM detailing method based on the adoption of DFs.
Experiments on 86 BIM doors from the NBS BIM Library were conducted to test the
proposed methodology. According to the experimental results, the proposed automatic BIM
detailing can achieve a satisfactory performance for all classification regression tasks with the
facilitation of DFs. The results showed that DFs contributed an average of 47.4% (up to
58.7%) of error reduction in the classification tasks and an average of 18.8% (up to 33.6%) in
the regression tasks. According to the performance and time cost in the experimental results,
SqueezeNet was the most cost-effective deep feature extraction architecture for the proposed
method.
This paper contributes to the research in two ways. From the industrial practitioner’s
perspective, the presented method demonstrates an automatic approach to BIM detailing.
BIM modelers can thus focus on the general designs, while the detailed BIM components –
even as invisible as hinges in the study case – can be created automatically. From BIM
researchers’ perspectives, the DFs of BIM 3D views and ML algorithms open new avenues
for BIM semantic enrichment and digital twin building. With a higher LOD of BIM, digital
twin constructions become more realistic and applicable for a variety of analyses.
Several research directions can extend this study. First, industrial specifications and standards
and manufacturing semantics can be incorporated to reinforce the BIM detailing process.
Furthermore, the non-geometry semantics in BIM can be further investigated if the semantic
properties, such as the weight or material proprieties, are consistently available. Also, the
applications of high LOD BIM objects and projects can be explored, such as interactive
AECO educational games and digital twins.
Acknowledgment
The work present in this paper was financially supported by Hong Kong Research Grants
Council (RGC) (Nos. 27200520, T22-504/21-R) and in part by The University of Hong Kong
(No. 202111159233).
References
Alhindi, T. J., Kalra, S., Ng, K. H., Afrin, A. & Tizhoosh, H. R. (2018). Comparing LBP,
HOG and deep features for classification of histopathology images. In Proceedings of

540

545

550

555

560

565

570

575

the 2018 International Joint Conference on Neural Networks. IEEE.
doi:10.1109/IJCNN.2018.8489329
Banihashemi, S., Khalili, S., Sheikhkhoshkar, M. & Fazeli, A. (2022). Machine learningintegrated 5D BIM informatics: Building materials costs data classification and
prototype development. Innovative Infrastructure Solutions, 7(3), 215.
doi:10.1007/s41062-022-00822-y
Barlish, K. & Sullivan, K. (2012). How to measure the benefits of BIM — A case study
approach. Automation in Construction, 24, pp. 149-159.
doi:10.1016/j.autcon.2012.02.008
Bianconi, F., Filippucci, M. & Buffi, A. (2019). Automated design and modeling for masscustomized housing. A web-based design space catalog for timber structures.
Automation in Construction, 103, pp. 13-25. doi:10.1016/j.autcon.2019.03.002
Boje, Y., Guerriero, A., Kubicki, S. & Rezgui, Y. (2020). Towards a semantic Construction
Digital Twin: Directions for future research. Automation in Construction, 114,
103179. doi:10.1016/j.autcon.2020.103179
Borrmann, A., König, M., Koch, C. & Beetz, J. (2018). Building Information Modeling.
Springer, ISBN 9783319928623. doi:10.1007/978-3-319-92862-3
Chan, D. W., Olawumi, T. O. & Ho, A. M. (2019). Perceived benefits of and barriers to
Building Information Modelling (BIM) implementation in construction: The case of
Hong Kong. Journal of Building Engineering, 25, 100764.
doi:10.1016/j.jobe.2019.100764
Chen, J., Li, S., Liu, D. & Lu, W. (2022). Indoor camera pose estimation via style‐transfer 3D
models. Computer-aided Civil and Infrastructure Engineering, 37(3), pp. 335-353.
doi:10.1111/mice.12714
Chen, K., Lu, W., Peng, Y., Rowlinson, S. & Huang, G. Q. (2015). Bridging BIM and
building: From a literature review to an integrated conceptual framework.
International Journal of Project Management, 33(6), pp. 1405-1416.
doi:10.1016/j.ijproman.2015.03.006
Chen, T. & Guestrin, C. (2016). XGBoost: A scalable tree boosting system. In Proceedings of
the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data
Mining (pp. 785-794). New York, NY, US: ACM. doi:10.1145/2939672.2939785
CIC. (2020). CIC BIM Standards - General (Version 2). Hong Kong, China: Construction
Industry Council. Retrieved June 21, 2022, from
https://www.bim.cic.hk/en/resources/publications_detail/85
CIC. (2021). Construction Digitalisation Roadmap for Hong Kong. Hong Kong, China:
Construction Industry Council. Retrieved January 25, 2023, from
https://www.bim.cic.hk/en/resources/publications_detail/95
Dargan, S., Kumar, M., Ayyagari, M. R. & Kumar, G. (2020). A survey of deep learning and
its applications: a new paradigm to machine learning. Archives of Computational
Methods in Engineering, 27(4), pp. 1071-1092. doi:10.1007/s11831-019-09344-w

580

585

590

595

600

605

610

615

Dong, N., Zhao, L., Wu, C.-H. & Chang, J.-F. (2020). Inception v3 based cervical cell
classification combined with artificially extracted features. Applied Soft Computing,
93, 106311. doi:10.1016/j.asoc.2020.106311
Fai, S. & Rafeiro, J. (2014). Establishing an appropriate level of detail (LoD) for a building
information model (BIM) – West Block, Parliament Hill, Ottawa, Canada. ISPRS
Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences, II5, pp. 123-130. doi:10.5194/isprsannals-II-5-123-2014
Ghaffarianhoseini, A., Tookey, J., Ghaffarianhoseini, A., Naismith, N., Azhar, S., Efimova,
O. & Raahemifar, K. (2017). Building Information Modelling (BIM) uptake: Clear
benefits, understanding its implementation, risks and challenges. Renewable and
Sustainable Energy Reviews, 75, pp. 1046-1053. doi:10.1016/j.rser.2016.11.083
Hall, M., Frank, E., Holmes, G., Pfahringer, B., Reutemann, P. & Witten, I. H. (2009). The
WEKA data mining software: An update. Association for Computing Machinery
Special Interest Group on Knowledge Discovery in Data Explorations Newsletter,
11(1), pp. 10-18. doi:10.1145/1656274.1656278
Han, J., Kamber, M. & Pei, J. (2011). Data Mining: Concepts and Techniques (3rd ed.).
Cambridge, MA, US: Morgan Kaufmann, ISBN 978-0-12-381479-1.
doi:10.1016/C2009-0-61819-5
Han, X.-F., Laga, H. & Bennamoun, M. (2019). Image-based 3D object reconstruction: Stateof-the-art and trends in the deep learning era. IEEE Transactions on Pattern Analysis
and Machine Intelligence, 43(5), pp. 1578-1604. doi:10.1109/TPAMI.2019.2954885
Hearst, M. A., Dumais, S. T., Osuna, E., Platt, J. & Scholkopf, B. (1998). Support vector
machines. IEEE Intelligent Systems and their Applications, 13(4), pp. 18-28.
doi:10.1109/5254.708428
Heaton, J., Parlikad, A. K. & Schooling, J. (2019). Design and development of BIM models
to support operations and maintenance. Computers in Industry, 111, pp. 172-186.
doi:10.1016/j.compind.2019.08.001
Hong, Y., Hammad, A. & Akbarnezhad, A. (2019). Forecasting the net costs to organisations
of Building Information Modelling (BIM) implementation at different levels of
development (LOD). Journal of Information Technology in Construction, 24, pp. 588603. doi:10.36680/j.itcon.2019.033
Hu, Y. & Castro-Lacouture, D. (2019). Clash Relevance Prediction Based on Machine
Learning. Journal of Computing in Civil Engineering, 33(2), 4018060.
doi:10.1061/(ASCE)CP.1943-5487.0000810
ISO. (2018). ISO 19650-1:2018 - Organization and Digitization of Information about
Buildings and Civil Engineering Works, including Building Information Modelling
(BIM) — Information Management using Building Information Modelling — Part 1:
Concepts and Principles. Geneva, Switzerland: ISO, ISBN 9789267110127.
Retrieved January 25, 2023, from https://www.iso.org/standard/68078.html

620

625

630

635

640

645

650

655

Kensek, K. (2015). Visual programming for building information modeling: Energy and
shading analysis case studies. Journal of Green Building, 10(4), pp. 28-43.
doi:10.3992/jgb.10.4.28
Kim, J. B., Jeong, W., Clayton, M. J., Haberl, J. S. & Yan, W. (2015). Developing a physical
BIM library for building thermal energy simulation. Automation in Construction, 50,
pp. 16-28. doi:10.1016/j.autcon.2014.10.011
LeCun, Y., Bengio, Y. & Hinton, G. (2015). Deep learning. Nature, 521(7553), pp. 436-444.
doi:10.1038/nature14539
Lee, H. J., Ullah, I., Wan, W., Gao, Y. & Fang, Z. (2019). Real-time vehicle make and model
recognition with the residual SqueezeNet architecture. Sensors, 19(5), 982.
doi:10.3390/s19050982
Lomio, F., Farinha, R., Laasonen, M. & Huttunen, H. (2018). Classification of Building
Information Model (BIM) structures with deep learning. In Proceedings of the 2018
7th European Workshop on Visual Information Processing. IEEE.
doi:10.1109/EUVIP.2018.8611701
Lou, J., Lu, W. & Xue, F. (2020). A review of BIM data exchange method in BIM
collaboration. In Proceedings of the 25th International Symposium on Advancement
of Construction Management and Real Estate (pp. 1329-1338). Springer Nature.
doi:10.1007/978-981-16-3587-8_90
Lu, Q., Parlikad, A. K., Woodall, P., Don Ranasinghe, G., Xie, X., Liang, Z., Konstantinou,
E., Heaton, J. & Schooling, J. (2020). Developing a digital twin at building and city
levels: Case study of West Cambridge campus. Journal of Management in
Engineering, 36(3), 5020004. doi:10.1061/(ASCE)ME.1943-5479.0000763
Natanian, J., Aleksandrowicz, O. & Auer, T. (2019). A parametric approach to optimizing
urban form, energy balance and environmental quality: The case of Mediterranean
districts. Applied Energy, 254, 113637. doi:10.1016/j.apenergy.2019.113637
NBS. (2022). NBS National BIM Library. Newcastle, UK: National Building Specification.
Retrieved July 7, 2022, from https://www.nationalbimlibrary.com
NIBS. (2015). National BIM Standard - United States Version 3. Washington, DC, US:
National Institute of Building Sciences. Retrieved June 10, 2022, from
https://www.nationalbimstandard.org/files/NBIMSUS_V3_3_Terms_and_Definitions.pdf
NIC. (2017). Data for the Public Good. London, UK: National Infrastructure Commission.
Retrieved July 16, 2022, from https://nic.org.uk/app/uploads/Data-for-the-PublicGood-NIC-Report.pdf
Opoku, D.-G. J., Perera, S., Osei-Kyei, R. & Rashidi, M. (2021). Digital twin application in
the construction industry: A literature review. Journal of Building Engineering, 40,
102726. doi:10.1016/j.jobe.2021.102726
Potseluyko, L., Pour Rahimian, F., Dawood, N., Elghaish, F. & Hajirasouli, A. (2022).
Game-like interactive environment using BIM-based virtual reality for the timber

660

665

670

675

680

685

690

695

frame self-build housing sector. Automation in Construction, 142, 104496.
doi:10.1016/j.autcon.2022.104496
Sadeghi, M., Elliott, J. W., Porro, N. & Strong, K. (2019). Developing building information
models (BIM) for building handover, operation and maintenance. Journal of Facilities
Management, 17(3), pp. 301-316. doi:10.1108/JFM-04-2018-0029
Seaton, H., Cristina, S., Sepasgozar, S. M. & Sawhney, A. (2022). Digital Twins from Design
to Handover of Constructed Assets. London, UK: Royal Institution of Chartered
Surveyors, ISBN 9781783214648. Retrieved Jul 18, 2022, from
https://www.rics.org/globalassets/wbef-website/reports-and-research/digital-twinsfrom-design-to-handover-of-constructed-assets.pdf
Song, J., Kim, J. & Lee, J. K. (2018). NLP and deep learning-based analysis of building
regulations to support automated rule checking system. In Proceedings of the 35th
International Symposium on Automation and Robotics in Construction. IAARC.
doi:10.22260/ISARC2018/0080
Su, T., Li, H. & An, Y. (2021). A BIM and machine learning integration framework for
automated property valuation. Journal of Building Engineering, 44, 102636.
doi:10.1016/j.jobe.2021.102636
Tang, S., Shelden, D. R., Eastman, C. M., Pishdad-Bozorgi, P. & Gao, X. (2019). A review of
building information modeling (BIM) and the internet of things (IoT) devices
integration: Present status and future trends. Automation in Construction, 101, pp.
127-139. doi:10.1016/j.autcon.2019.01.020
Thabet, W., Lucas, J. & Srinivasan, S. (2022). Linking life cycle BIM data to a facility
management system using Revit Dynamo. Organization, Technology & Management
in Construction, 14(1), pp. 2539-2558. doi:10.2478/otmcj-2022-0001
Torres-Calderon, W., Chi, Y., Amer, F. & Golparvar-Fard, M. (2019). Automated mining of
construction schedules for easy and quick assembly of 4D BIM simulations. In
Computing in Civil Engineering 2019: Visualization, Information Modeling, and
Simulation (pp. 432-438). ASCE. doi:10.1061/9780784482421.055
Valero, E., Forster, A., Bosché, F., Renier, C., Hyslop, E. & Wilson, L. (2018). High level-ofdetail BIM and machine learning for automated masonry wall defect surveying. In
Proceedings of the 35th International Symposium on Automation and Robotics in
Construction. IAARC. doi:10.22260/ISARC2018/0101
Wang, H., Pan, Y. & Luo, X. (2019). Integration of BIM and GIS in sustainable built
environment: A review and bibliometric analysis. Automation in Construction, 103,
pp. 41-52. doi:10.1016/j.autcon.2019.03.005
Weng, Y., Mohamed, N. A., Lee, B. J., Gan, N. J., Li, M., Jen Tan, M., Li, H. & Qian, S.
(2021). Extracting BIM information for lattice toolpath planning in digital concrete
printing with developed dynamo script: A case study. Journal of Computing in Civil
Engineering, 35(3), 5021001. doi:10.1061/(ASCE)CP.1943-5487.0000964

700

705

710

715

720

725

730

Wu, Y., Shang, J. & Xue, F. (2021). RegARD: Symmetry-based coarse registration of
smartphone’s colorful point clouds with CAD drawings for low-cost digital twin
buildings. Remote Sensing, 13(10), 1882. doi:10.3390/rs13101882
Xue, F. & Lu, W. (2020). A semantic differential transaction approach to minimizing
information redundancy for BIM and blockchain integration. Automation in
Construction, 118, 103270. doi:10.1016/j.autcon.2020.103270
Xue, F. (2022). As-built building information modeling: joint effort of 3D reconstruction and
semantic enrichment. In Research Companion to Building Information Modeling (pp.
190-211). Northampton, MA, USA: Edward Elgar Publishing.
doi:10.4337/9781839105524.00019
Xue, F., Lu, W., Webster, C. J. & Chen, K. (2019). A derivative-free optimization-based
approach for detecting architectural symmetries from 3D point clouds. ISPRS Journal
of Photogrammetry and Remote Sensing, 148, pp. 32-40.
doi:10.1016/j.isprsjprs.2018.12.005
Xue, F., Wu, L. & Lu, W. (2021). Semantic enrichment of building and city information
models: A ten-year review. Advanced Engineering Informatics, 47, 101245.
doi:10.1016/j.aei.2020.101245
Yang, L., Cheng, J. C. & Wang, Q. (2020). Semi-automated generation of parametric BIM
for steel structures based on terrestrial laser scanning data. Automation in
Construction, 112, 103037. doi:10.1016/j.autcon.2019.103037
Yang, Z., Xue, F. & Lu, W. (2021). Handling missing data for construction waste
management: machine learning based on aggregated waste generation behaviors.
Resources, Conservation and Recycling, 175, 105809.
doi:10.1016/j.resconrec.2021.105809
Zabin, A., González, V. A., Zou, Y. & Amor, R. (2022). Applications of machine learning to
BIM: A systematic literature review. Advanced Engineering Informatics, 51, 101474.
doi:10.1016/j.aei.2021.101474
Zhang, M., Ke, J., Tong, L. & Luo, X. (2021). Investigating the influence of route turning
angle on compliance behaviors and evacuation performance in a virtual-reality-based
experiment. Advanced Engineering Informatics, 48, 101259.
doi:10.1016/j.aei.2021.101259
Zhang, M.-L. & Zhou, Z.-H. (2007). ML-KNN: A lazy learning approach to multi-label
learning. Pattern Recognition, 40(7), pp. 2038-2048.
doi:10.1016/j.patcog.2006.12.019
Zhao, R., Chen, Z. & Xue, F. (2023). A blockchain 3.0 paradigm for digital twins in
construction project management. Automation in Construction, 145, 104645.
doi:10.1016/j.autcon.2022.104645