OPEN ACCESS

Vietnam Journal of Computer Science

Vol. 11, No. 1 (2024) 1–22
.c The Author(s)
#
DOI: 10.1142/S2196888823300028

Machine Learning Methods in BIM-Based

Applications | A Review
by 58.84.143.183 on 06/01/24. Re-use and distribution is strictly not permitted, except for Open Access articles.

Grażyna Ślusarczyk * and Barbara Strug †

Institute of Applied Computer Science, Jagiellonian University

Vietnam J. Comp. Sci. 2024.11:1-22. Downloaded from worldscientific.com

Lojasiewicza 11, 30-059 Krak
ow, Poland

*[email protected]
†[email protected]

Received 26 July 2023

Revised 30 September 2023
Accepted 1 October 2023
Published 4 November 2023

This paper presents a survey of machine learning (ML) methods used in applications dedicated
to the building and construction industry. A building information modeling (BIM) model, being
a database system for civil engineering data, is presented. A representative selection of methods
and applications is described. The aim of this paper is to facilitate the continuation of research
e®orts and to encourage bigger participation of database system researchers in the ¯eld of civil
engineering.

Keywords: BIM data; machine learning; civil engineering modeling.

1. Introduction
Building Information Modeling (BIM) is nowadays widely used in architecture, en-
gineering and construction industry (AEC). The building and construction industry
employs currently about 7% of the world's working-age population and is one of the
world economy's largest sectors. It is estimated that about $10 trillion is spent on
construction-related goods and services every year. In the last decade, the acceptance
and actual use of BIM has increased signi¯cantly within the building community.
It has largely contributed to the process of eliminating faults in designs. BIM allows
architects and engineers to create 3D simulations of the desired structures which
contain signi¯cantly more information on the actual structures than drawings pro-
duced using traditional Computer Aided Drafting (CAD) systems. As a result, BIM
has become more and more present in the construction industry.

† Corresponding author.
This is an Open Access article published by World Scienti¯c Publishing Company. It is distributed under
the terms of the Creative Commons Attribution 4.0 (CC BY) License which permits use, distribution and
reproduction in any medium, provided the original work is properly cited.

1
 2

G. Slusarczyk & B. Strug

BIM technology enables representation of syntactic and semantic building

information with respect to the entire life cycle of designed objects, from the design
phase, through construction to the facility management (FM) phase. BIM includes
information about the elements and spaces within buildings, their constituting ele-
ments, their interrelations, properties and performance. The project created in BIM
technology can be treated as a database that allows to record both technical infor-
mation about building elements, and its purpose and history. However, although
by 58.84.143.183 on 06/01/24. Re-use and distribution is strictly not permitted, except for Open Access articles.

BIM is information rich, not all knowledge is explicitly stated. It seems that ML
approaches suit well to deduce implicit knowledge from BIM models. Contrary to
querying approaches used to extract knowledge from building models,1–3 which are
tailored to speci¯c scenarios with prede¯ned outcomes, ML methods are able to
Vietnam J. Comp. Sci. 2024.11:1-22. Downloaded from worldscientific.com

detect patterns and make predictions.

Using ML and arti¯cial intelligence (AI) in AEC industry is a promising research
direction. It has to be noted that while both ML and AI are rapidly developing across
many other industries, the construction industry is lagging behind in the rate by
which improvements are introduced. The usage of BIM could be seen as a tool to
revert this trend, but software tools able to implement BIM still require quite la-
borious routine tasks to properly execute BIM.
The structure of data in BIM, where knowledge is represented in an object ori-
ented way, is ideal for analytical purposes and the application of ML techniques.4 ML
is related to extrapolating object behaviors and generating logical responses from
information provided by examples, and enabling a computer to gradually learn.
Various classi¯cation algorithms, anomaly detection, and time series analysis can be
used with respect to BIM. Classi¯cation algorithms can be used for example to
predict characteristics of °ats and their sale demand, the likelihood of construction
delays, or to diagnose the assets in historic buildings. Anomaly detection is useful in
assessment of architectural models, discovering modeling errors, while time series
analysis can be applied to make maintenance predictions or renovations planning.
In order to make predictions or detect patterns, metrics about buildings are to be
speci¯ed. They serve as a label to building models which allows for measuring their
performance.
This paper is an extended version of the conference paper (see Ref. 5). A signif-
icant number of new papers have been carefully selected for this extension. The new
paper both adds deeper analysis of previously presented domains of applications
and researches other applications within the BIM methodology.
The objective of this paper is to present highlights of references pertaining to ML
in BIM. It complements previously published literature survey articles in order to
provide insight into the development of AI in BIM, underlines the hotspots of current
research in this domain, and facilitates continued research e®orts. The paper
summarizes recently developed theories and methods applied in BIM-based knowl-
edge processing, extraction, and semantic enrichment of BIM models. They include
neural networks, decision trees, logistic regression, a±nity propagation clustering,
 ML Methods in BIM-Based Applications 3

term frequency (TF), random forests, SVM, as well as Bayesian networks. The paper
provides an overview of the advances of ML methods applied in BIM.

2. BIM Data | IFC Files

The information about a building created in any software can be exported to dif-
ferent formats. Each commercial application has its own ¯le type to store building
by 58.84.143.183 on 06/01/24. Re-use and distribution is strictly not permitted, except for Open Access articles.

data, but all of them can also export building information to an Industry Foundation
Classes (IFC) ¯le. The ¯le format IFC has become the de facto standard way of
interchanging and storing BIM data.6 It is an interoperable BIM standard for CAD
applications, which supports a full range of data exchange among di®erent disciplines
Vietnam J. Comp. Sci. 2024.11:1-22. Downloaded from worldscientific.com

and heterogeneous applications. Information retrieved from IFC ¯les can be used by
many di®erent applications.
IFC speci¯es di®erent types of building entities and their basic properties.
It de¯nes an EXPRESS-based entity-relationship model, which consists of several
hundred entities organized into an object-based inheritance hierarchy. All the enti-
ties in IFC are divided into rooted and non-rooted ones. While the ¯rst ones are
derived from IfcRoot and have identity (a GUID), attributes for name, description,
and revision control, the other ones (non-rooted) do not have an identity and their
instances exist only if they are referenced from a rooted instance directly or indi-
rectly. IfcRoot is subdivided into three concepts: Object de¯nitions, relationships,
and property sets:

. IfcObjectDe¯nition captures tangible object occurrences and types.

. IfcRelationship captures relationships among objects.
. IfcPropertyDe¯nition captures dynamically extensible properties of objects.

In Fig. 1 a fragment of an example visualization of an IFC ¯le representing a multi-

storey building is depicted. In the top left-hand side panel, the hierarchical structure
of the ¯le is shown. The °oor elements describe the storeys of the building. One of the
°oor elements is developed to show its di®erent component entities of the type
IfcBeam, IfcColumn and IfcWall in a more detailed way. The component entities of
one of the IfcWall elements are shown and their visualization is depicted by the
darker color in the right-hand side panel in Fig. 1.
The entities that can be used in IFC include building components like IfcWall,
IfcDoor, IfcWindow, geometry such as IfcExtrudedAreaSolid, and basic constructs
such as IfcCartesianPoint.
The most often used building elements are IfcSpace, IfcDoor, IfcWall, IfcStair,
and IfcWindow. According to the IFC 2x Edition3 Model Implementation Guide6
and the IFC speci¯cation6 the above-mentioned classes can be described as follows:

(1) IfcSpace is the instance used to represent a space as the area or volume of a
functional region. It is often associated with the class IfcBuildingStorey repre-
senting one °oor (the building itself is an \aggregation" of several storeys) or
 4

G. Slusarczyk & B. Strug
by 58.84.143.183 on 06/01/24. Re-use and distribution is strictly not permitted, except for Open Access articles.
Vietnam J. Comp. Sci. 2024.11:1-22. Downloaded from worldscientific.com

Fig. 1. An example of IFC data visualization.

with IfcSite, which represents the construction site. A space in the building is
usually associated with certain functions (e.g. kitchen or bathroom). These
functions are speci¯ed by attributes of the class IfcSpace (Name, LongName,
Description).
(2) IfcWall is the instance used to represent a vertical element, which is to merge or
split the space. In IFC ¯les two representations of a wall can be distinguished.
The subclass IfcWallStandardCase of IfcWall is used for all walls that do not
change their thickness (the thickness of a wall is the sum of the materials used).
IfcWall is used for all other walls, in particular for constructs of varying thickness
and for the walls with non-rectangular cross-sections.
(3) IfcStair represents a vertical passage allowing for moving from one °oor to the
other. It can contain an intermediate landing. Instances of IfcStairs are treated as
containers, by which we refer to component elements as IfcStairFlight using
IfcRelAggregates.
(4) IfcDoor represents a building element used to provide access to a speci¯c area or
room. Parameters of IfcDoors specify dimensions, an opening direction and a
style of the door (IfcDoorStyle). IfcDoor is a subclass of IfcBuildingElement and
a superclass of IfcDoorStandardCase. Door instances are usually located in a
space IfcOpeningElement to which we refer by IfcRelFillsElement.
(5) IfcWindow represents a building element used to ¯ll vertical or near-vertical
openings in walls or roofs. It provides a view, light and fresh air. The dimensions
of the window and its shape can be found in IfcWindowStyle, to which we refer
by IfcRelDe¯nesByType. IfcWindow is a subclass of IfcBuildingElement and a
superclass of IfcWindowStandardCase. Window instances are placed in a space
IfcOpeningElement to which we refer by IfcRelFillsElement.
 ML Methods in BIM-Based Applications 5
by 58.84.143.183 on 06/01/24. Re-use and distribution is strictly not permitted, except for Open Access articles.
Vietnam J. Comp. Sci. 2024.11:1-22. Downloaded from worldscientific.com

Fig. 2. IfcWall and IfcDoor within IFC hierarchy.

The above-mentioned instances inherit from IfcProduct class, which allows for de-
termining their positions using geometrical entities like IfcLocalPlacement and
relationships like IfcPlacementRelTo. The place of IfcWall and IfcDoor within the
IFC hierarchy is depicted in Fig. 2 (from Ref. 8).

3. ML Techniques for BIM

Many recent studies address the need for semantic integrity in BIM models as it
allows for sharing these models between project teams, performing domain-speci¯c
analyses on BIM model data and checking BIM model deliverables by owners.
Moreover, information important from the point of view of many applications lies
implicit in the interrelation between building elements. Therefore, several approa-
ches directed at extracting implicit data from building models have been presented.
They explore mainly an inductive reasoning approach, which uses ML and deep
learning, where rules are learned and generalized from the data, to ensure the in-
tegrity of BIM element classi¯cations. Bloch and Sacks showed that a ML-based
approach is more successful in automatically classifying spaces than deductive
reasoning, which is based on a rule-inferencing approach.9

3.1. Learning semantic information

3.1.1. Space classi¯cation
An unsupervised learning method for mining the IFC-based BIM data by exploring
interrelations between building spaces is presented in Ref. 10. A method of extracting
features, which are then used in the a±nity propagation clustering algorithm to get
spaces with similar usage functions, is proposed. The method allows for automatic
learning of functional knowledge from building space structures. The physical
properties of each space and their boundary relationships in a BIM model are
 6

G. Slusarczyk & B. Strug

extracted from the IFC ¯le based on BIM data. Then boundary graphs with space
boundary relationships, where properties of each space propagate along the edges,
are built. Features of building spaces are extracted from the space boundary graphs.
Based on these features and the graph representation of the building structure, the
adapted a±nity propagation algorithm performs building space clustering analysis, in
order to get representative samples of building spaces. The experimental results per-
formed on a real-world BIM dataset containing 595 spaces from a 20-storey building
by 58.84.143.183 on 06/01/24. Re-use and distribution is strictly not permitted, except for Open Access articles.

show that building spaces with typical usage functions, like senior o±ces, open o±ces
and circulation spaces, can be discovered by the unsupervised learning algorithm.
In Ref. 4, rooms in a housing unit were named according to their use (dining
room/lounge, kitchen, bedroom, etc.) based on their geometry. The di®erent types of
Vietnam J. Comp. Sci. 2024.11:1-22. Downloaded from worldscientific.com

rooms in BIM are usually labeled entirely by hand by the expert designer. Using ML
algorithms to automate this type of task considerably reduces the computational
time. Three di®erent classi¯cation algorithms, namely decision trees, logistic
regressions, and neural networks, were used to solve the problem of labeling rooms
according to their function. The input data from which the algorithms are to learn
consist of rooms in the housing unit whose function was labeled by hand previously.
The data were obtained from two models of housing developments created with
Autodesk Revit, each consisting of more than 200 housing units, from which some
rooms were extracted. The algorithms were trained with one project and evaluated
with another, which ensures that the results could be extrapolated to other projects.
In order to compare the sensitivity of the amount of information available, two
di®erent datasets were created for each model. The ¯rst model only includes the
information obtained directly (or via formulae) from the Revit schedules. The second
model also includes information that can only be extracted or calculated by C#
programming using the Revit API. Predictions made by logistic regression or neural
network with complete data were about 80–90% accurate in predicting room use.
The most common errors are mistaking kitchens for bedrooms and classifying cor-
ridors or bedrooms as bathrooms.
Experimental approaches directed at extracting implicit data from building
models were also researched.11 Both unsupervised and supervised ML of BIM models
are considered. A supervisor ML approach, which is based on a neural network, is
able to classify °oor plans according to their intended function. By looking at the
spatial con¯gurations of °oor plans, a neural network was trained to di®erentiate
between residential and institutional facilities. This approach can be used to com-
plete missing attributes in datasets, where information pertaining to the intended
function and use is fragmented and incomplete. Both supervised and unsupervised
learning algorithms assess a building by means of a set of its characteristic features.
The IFC ML platform presented in this paper is built on top of the DURAARK IFC
metadata extractor.12 This tool is able to extract literal values, aggregate and derive
values from IFC SPF ¯les. In Ref. 13, a dataset of IFC ¯les has been proposed to
facilitate the comparison of the classi¯cation results for di®erent IFC entities. The
unsupervised learning approach is used in Ref. 11 to show anomalies in building
 ML Methods in BIM-Based Applications 7

models. The obtained results make it possible to °ag uncommon situations (like
unusual large overhangs or an unusual con°uence of several building elements) that
might need additional checks or coordination, and therefore reduce the failure costs
in the construction industry.
In Ref. 9 ML algorithms are used to the problem of classi¯cation room types in
residential apartments. Classi¯cation of room types and space labeling are important
for the design process, compliance checking, management operations, and many
by 58.84.143.183 on 06/01/24. Re-use and distribution is strictly not permitted, except for Open Access articles.

building analysis tasks. The dataset used for supervised ML processes contains
32 BIM apartment models. The classi¯cation and labeling of room types in this work
is based on their function and assumes that spaces do not have dual function.
A multiclass neural network was used with a total of 150 spaces in the dataset for the
Vietnam J. Comp. Sci. 2024.11:1-22. Downloaded from worldscientific.com

training process. The dataset was split into 70% for training and 30% for validation
of the trained model, which resulted in an 82% correctly classi¯ed validation set. The
building objects were classi¯ed based on the ¯ve local features, area, number of
doors, number of windows, number of room boundary lines, and °oor level o®set, and
one connecting feature being direct access. The obtained results showed that ML is
directly applicable to the space classi¯cation problem.
Four di®erent ML methods are used to categorize images extracted from BIM of
building designs.14 BIM data are separated into three categories: apartment build-
ings, industrial buildings and others. The ¯rst method is based on classical ML,
where Histogram of Oriented Gradients (HOG) is used to extract features, and a
Support Vector Machine (SVM) is used for classi¯cation. The other three methods
are based on deep learning. The ¯rst two use pre-trained Convolutional Neural
Networks (CNN) (a MobileNet15 and a Residual Network16). The third one is a CNN
with a randomly generated structure. A database of 240 images extracted from
60 BIM virtual representations is used to validate the classi¯cation precision of the
models. The accuracy achieved by the HOG+SVM model is 57%, while for the neural
networks it is above 89%. The approach shows that it is possible to automatically
categorize a structure type from a BIM representation.

3.1.2. Object classi¯cation

In Ref. 17 a deep learning approach was proposed to classify 3D models in BIM
environments. A ray-based feature extraction algorithm was used to extract the fea-
tures of 3D models and form a feature matrix. This matrix is applied by the deep belief
network (DBN) constructed by restricted Boltzmann machines, which classi¯es the
models. The database included 1814 models, covering most types of models common in
daily life, which were divided into a training set and a testing set, each with 907 models.
The results show that the proposed approach achieves accuracy of 85.5%.
Koo et al. applied the SVM technique to classify building elements for checking
the semantic integrity of mappings between BIM elements and IFC classes.18 The
SVM was trained to distinguish model elements from a dataset of 4187 unique
elements collected from six architectural BIM models, based on their geometric and
 8

G. Slusarczyk & B. Strug

relational features. In the ¯rst stage, the SVM was tested to classify the elements
with respect to eight IFC classes (wall, slab, beam, door, window, railing, covering,
and column). The dataset was divided into training and test sets by a ratio of eight to
two resulting in a test set of 838 elements on which 10-fold cross-validation was
performed. In the second stage, the SVM was further tested to distinguish between
the element subtypes within three IFC classes, IfcDoor (single and double door),
IfcWindow (single and double window), and IfcColumn (rectangular and non-rect-
by 58.84.143.183 on 06/01/24. Re-use and distribution is strictly not permitted, except for Open Access articles.

angular shaped column). The same training-to-test ratio was used and 10-fold cross-
validation was performed to optimize hyperparameters. The approach showed
94.39% accuracy of classifying building elements in the ¯rst stage, while 100% for the
window and column types, and 96% for door types in the second stage.
Vietnam J. Comp. Sci. 2024.11:1-22. Downloaded from worldscientific.com

As it is di±cult to ensure the BIM data integrity due to errors, mistakes, or

omissions of object information, Kim et al. used 2D CNN to recognize and classify
unknown objects as BIM instances.19 The presented approach classi¯es basic
architectural BIM elements, as well as furniture entities according to their images
using a 3D object recognition technique. In the ¯rst stage the class of an object is
recognized (slab, wall, ceiling, column, staircase, door, seating furniture, and toilet
¯xture), while in the second one, subtypes of door, seating furniture and toilet ¯xture
are covered. The dataset consists of 820 objects spanning 10 classes. The model input
consists of multi-view 2D rendering images of BIM objects. The training of the
recognition models is processed by ¯ne-tuning the Inception-v3 model20 that trains
the ImageNet21 dataset on TensorFlow framework.22 80% of the data is used for
training and 20% for validation. The accuracy of training for the classi¯cation of
building elements is 96%. The second stage training results showed an average of
96.7% accuracy of classi¯cation of sub-types of doors, seating and toilet ¯xture.
3D geometric deep neural networks were used for automatic classi¯cation of BIM
elements by Koo et al.23 Multi-view CNN (MVCNN)24 and PointNet25 were inves-
tigated to determine their applicability in extracting unique features of door (Ifc-
Door) and wall (IfcWall) element subtypes. MVCNN builds classi¯ers of 3D shapes
from 2D image renderings of these shapes, while PointNet is a deep neural network,
which takes point clouds directly as input to classify and segment 3D artifacts.
Firstly 478 door elements and 364 wall elements were collected and divided into their
respective subtypes as single, double, sliding and revolving door, and a generic wall,
wall with window opening, and wall with door opening. The samples were divided
using a 7:3 ratio to train and test the learning models. Test results indicated
MVCNN as having the best prediction performance with accuracy of 92% and 95%
for door and wall element subtypes, respectively, while PointNet's accuracy was
hampered by resolution loss due to selective use of point cloud data.
Geometric deep learning models were also used to classify infrastructure BIM
elements for checking the integrity of element-to-entity mappings in BIM models.26
MVCNN and PointNet models were trained and tested to classify 10 types of
commonly used BIM elements in road infrastructure. They include columns, three
types of culverts, four types of retaining walls, sumps, and wing walls. A dataset
 ML Methods in BIM-Based Applications 9

consists of 1496 3D elements, which were divided into training and test sets, using a
7:3 ratio. Results revealed MVCNN as the superior model with an accuracy of 98%,
compared with PointNet's corresponding value of 83%.
Collins et al. encoded BIM objects using two kinds of graph encodings and utilized
a graph convolutional network (GCN) to perform classi¯cation based on the 3D
shape only.27 A GCN uses the native triangle-mesh and automatically creates
meaningful local features for subsequent classi¯cation. The data are extracted from
by 58.84.143.183 on 06/01/24. Re-use and distribution is strictly not permitted, except for Open Access articles.

22 IFC tagged ¯les which are divided into training and test sets, using an 8:2 ratio.
The dataset consists of structural elements (IfcWall, IfcSlab, IfcColumn, IfcWindow,
IfcDoor, IfcStair IfcRailing), the equipment (IfcFlowTerminal, IfcFlowSegment,
IfcFlowFitting, IfcDistributionControlElement, IfcFlowController) and the interior
Vietnam J. Comp. Sci. 2024.11:1-22. Downloaded from worldscientific.com

furniture (IfcFurnishingElements). The method reaches an accuracy of up to 85% on

the dataset collected by authors.
Luo et al. proposed a geometric-relational deep learning framework for BIM ob-
ject classi¯cation,28 which boosts the classi¯cation methods based on object geom-
etry with relational information. The proposed framework consists of three main
modules, i.e. a geometric feature extractor, relational feature extractor and feature
fusion module. The geometric feature extractor learns to represent a BIM object's
shape as a geometric descriptor. The relational feature extractor extracts relational
features of di®erent levels and connects them to form the relational descriptor. The
feature fusion module mixes the two descriptors and outputs the ¯nal object de-
scriptor. Three proposed relational models RMVCNN, RDGCNN, RMVViT are
based on MVCNN, DGCNN29 and MVViT (a 3D deep learning model adapted from
Vision Transformer30), respectively, which act as geometric feature extractors. The
IFCNet++ dataset, where selected representative relational features are attached to
each BIM object, was proposed and used in all the experiments for training and
testing the described models. The dataset contains 9228 objects belonging to 11
types. Four subtypes of IfcRelationship that appear in most IFC ¯les (IfcRelCon-
nectsElement, IfcRelFillsElement, IfcRelAggregates, and IfcRelVoidsElement) were
selected. The data were divided into training and test sets using a 7:3 ratio. The
results show that RMVCNN reaches an accuracy higher than 98% on 9 of the 11
types, which means that it can e®ectively learn geometric and relational features of
most types. Accuracy on the other two models is slightly lower than 90%. Experi-
ments show that relational features are representative enough to help the geometric
methods achieve a performance gain.

3.2. Semantic enrichment of BIM models from point clouds

The other group of papers considers the use of ML algorithms for semantic enrich-
ment of BIM models obtained from point cloud data.31 Semantic enrichment tech-
niques solve the interoperability problem by exploiting existing numeric, geometric,
or relational information in the model to infer new semantic information. In this way,
the time-consuming process of manually creating 3D models useful for architectural
 10

G. Slusarczyk & B. Strug

and civil engineering applications can be avoided. Semantic enrichment encompasses

classi¯cation of building objects, aggregation and grouping of building elements,
implementing associations to re°ect connections and numbering,32 unique identi¯-
cation, completion of missing objects, and reconstruction of occluded objects. Then
the classi¯cation of the model as a whole, or of particular assemblies or objects within
the model with respect to code compliance, can be performed. BIM objects derive
many of their properties from their class, making object classi¯cation crucial for reuse
by 58.84.143.183 on 06/01/24. Re-use and distribution is strictly not permitted, except for Open Access articles.

in di®erent analysis tasks, like spatial validation of a BIM model, quantity take-o®
and cost estimation.
Xiong et al. used ML for classifying and labeling surfaces obtained from a laser
scan, in order to semantically enrich a BIM model.33 They used both the shape
Vietnam J. Comp. Sci. 2024.11:1-22. Downloaded from worldscientific.com

features, which are referred to as local features, and the spatial relationships
(orthogonal, parallel, adjacent, and coplanar) referred to as contextual features.
A context-based ML algorithm, called stacked learning,34 is used to label patches
extracted from a voxelized version of the input point cloud. The main constructive
objects of a building, i.e. walls, ceilings, and °oors, were classi¯ed, and separated
from other objects obtained from the scan that are considered clutter. The method
achieved an average 85% accuracy over four classes. Then the SVM algorithm is used
to encode the characteristics of an opening shape and location, which allows the
algorithm to estimate the shape of window and doorway openings even when they are
partially occluded. The method was evaluated on a large, highly cluttered dataset of
a two-storey schoolhouse building containing 40 rooms. The facility was scanned
from 225 locations resulting in over 3 billion 3D measurements. The SVM algorithm
was able to detect window and door openings with 88% accuracy.
In Ref. 35 Emunds et al. proposed an e±cient neural network to learn from point
cloud representation of BIM objects. Sparse Residual CNN for Semantic Enrichment
of BIM models (SpaRSE-BIM), a neural network model based on sparse convolu-
tions, is used for the classi¯cation of IFC-based geometry and semantic enrichment of
BIM models. Sparse convolutions only store the non-empty parts of the space and
compute outputs for prede¯ned coordinates, making them more e±cient and faster
than regular 3D convolutions on sparse data. The structure of SpaRSE-BIM is made
up of three building blocks, the Sparse Conv Block, the MLP Block, and the Sparse
Residual Block, which is the main building block of the network. SpaRSE-BIM
processes point clouds. The experiments were performed on the IFCNetCore and
BIMGEOM datasets. IFCNet provides a dataset of 19,613 individual IFC objects
from 65 IFC classes, its subset, called IFCNetCore, provides 7930 objects from 20
classes. BIMGEOM consists of 10,146 objects of structural elements, technical
equipment and interior furniture and contains 13 IFC classes. The results indicate
that SpaRSE-BIM achieves overall accuracy of 81.59% on IFCNetCore and 90.47%
on BIMGEOM.
The ML methods for both semantic enrichment and automating design review
using BIM models are proposed.36 The approach was applied to identify security
rooms and their walls within the BIM model created in the design process.
 ML Methods in BIM-Based Applications 11

A two-class decision forest classi¯cation algorithm was chosen for this experiment.37
It was implemented on a dataset of models with 642 security rooms with non-regular
complex geometry, arranged in 64 shafts. The dataset for training contained 448
spaces, 278 of which were security rooms compliant with the described code clause
and the remaining 170 were other rooms or open spaces. The spaces were organized in
64 vertical shafts, each of which was comprised of security rooms and other spaces.
Running a 10-fold cross-validation algorithm achieved 88% accuracy. The authors
by 58.84.143.183 on 06/01/24. Re-use and distribution is strictly not permitted, except for Open Access articles.

explored the possibility of developing a deep neural network to classify the rooms
when the only input data are a wall schedule, a room schedule and a table of the
relationships between them.
Vietnam J. Comp. Sci. 2024.11:1-22. Downloaded from worldscientific.com

3.3. Supporting building design

Di®erent types of data available in BIM models can be used to support the building
design phase in the AEC industry. They help to optimize the design, detect clashes,
determine the building energy consumption in order to improve sustainability of the
construction, better understand client's needs, and make better-informed decisions.
Chen et al. proposed a framework combining BIM with the least square support
vector machine (LSSVM) and non-dominated sorting genetic algorithm-II (NSGA-
II) to study the in°uence of building envelope parameters on building energy con-
sumption.38 The LSSVM is utilized to learn the dataset to establish a prediction
model between the building envelope and energy consumption, which is used as the
¯tness function of the building energy consumption in NSGA-II. An NSGA-II multi-
objective optimization model with the objective of minimizing building energy
consumption and maximizing indoor thermal comfort is established. The proposed
framework was used to optimize the design of a school building. The results show
that the LSSVM has great ability to predict building energy consumption, achieving
a prediction accuracy of 95.49%. The developed LSSVM-NSGA II hybrid approach
can e®ectively improve the building's energy consumption and thermal comfort with
the building energy consumption reduced by 10.6% and the thermal comfort in-
creased by 32.2% compared to the initial values, respectively. A uni¯ed framework
for the extraction and integration of BIM models and building energy consumption
monitoring data has been proposed.39 The energy usage anomaly diagnosis methods
have been devised using integrated static and dynamic data, which employ density-
based clustering and arti¯cial neural network (ANN) regression using the integrated
data. The multi-layered backpropagation neural network model regressing over
input data from a multitude of buildings, with building static data treated as input
parameters, is used to identify implicit anomalies that are beyond the scope of
clustering methods.
In Ref. 40 ML is used to predict building thermal energy performance. A com-
ponent-based approach that develops ML models for a parameterized whole building
design and for parameterized components of the design is described. ML components
have been developed at two levels of decomposition, namely construction-level
 12

G. Slusarczyk & B. Strug

components (wall, window, °oor, roof, etc.) with a static prediction of total yearly
heating and cooling energy consumption as a response, and zone-level components
with dynamic monthly consumption as a response. The development of ML com-
ponents representing construction elements uses neural networks with one hidden
layer per component. Training data were split into a training set (85%) and a set
used for independent cross-validation of components after testing (15%). Two
algorithms were used for training, namely the Levenberg–Marquardt algorithm and
by 58.84.143.183 on 06/01/24. Re-use and distribution is strictly not permitted, except for Open Access articles.

Bayesian regularization. Results in test cases show that high prediction quality is
achieved with errors of 3.7% for cooling and 3.9% for heating. In Ref. 41 the com-
ponent-based ML using ANNs is used for energy performance prediction in the early
stages of building design. As the application of BIM in early design phases requires
Vietnam J. Comp. Sci. 2024.11:1-22. Downloaded from worldscientific.com

the support of di®erent levels of detail, the approach links multiLOD building models
with ML components. This approach was illustrated and validated by a test case
with a medium-sized o±ce building. For validation purposes, the results of the ML
predictions for 60 di®erent design con¯gurations were considered, showing that the
average error was no higher than 3.5% for cooling and 3.8% for heating.
In Ref. 42 a system based on BIM and ANN, which evaluates indoor personal
thermal comfort and energy of the interior space design, is proposed. It integrates
thermal information and a building model. The backpropagation ANN predictive
model of thermal comfort takes three environment parameters (air temperature, air
humidity, and wind speed around the person), three human state parameters
(human metabolism rate, clothing thermal resistance, and the body position) and
four body parameters (gender, age, height, and weight) as inputs. The proposed
plugin of BIM provides thermal comfort evaluation results and energy saving opti-
mization suggestions for interior space design by simulating the energy consumption
index of di®erent design schemes. In the data test, the Mean Squared Error of the
established ANN model was equal to 0.39. In Ref. 43 a framework based on BIM and
ML data-driven models, with a goal to analyze the optimum thermal comfort for
indoor environments with the e®ect of natural ventilation, is proposed. Su±cient
natural ventilation in buildings is important for reducing the energy consumption of
mechanical ventilation while maintaining a comfortable indoor environment for
occupants. BIM provides geometrical and semantic information of the built envi-
ronment, which is needed for specifying boundary conditions of computational °uid
dynamics (CFD) simulation. This simulation is conducted to obtain the °ow ¯eld
and temperature distribution, the results of which determine the thermal comfort
index in a ventilated environment. BIM-CFD provides spatial data, boundary con-
ditions, indoor environmental parameters, and the thermal comfort index for ML to
construct robust data-driven models to empower the predictive analysis. In the
neural network, the spatial features such as zone adjacency and connectivity, and the
impact of air°ow between zones are represented in the form of the adjacency matrix.
A system, which aims to assist architects in designing buildings with minimal
energy costs while meeting the functional purpose of the designed facility, is proposed
in Ref. 44. It integrates BIM and ANN to predict energy cost of residential buildings
 ML Methods in BIM-Based Applications 13

at the conceptual stage of architectural design. BIM enables the designer to produce
di®erent scenarios of architectural design, while the ANN model is used to predict the
energy cost of the designs de¯ned in BIM. Datasets used for developing the predic-
tion model contain factors in°uencing the cost of energy such as the age of the
existing residential building, location, building area, number of occupants, glazing
type, insulation, a building envelope system and air conditioner type. One-hundred-
and-eighty-six datasets were used to build the model, 144 datasets were utilized for
by 58.84.143.183 on 06/01/24. Re-use and distribution is strictly not permitted, except for Open Access articles.

training, 16 datasets were used for testing the model and 26 datasets were used for
validation. The results show that the developed ANN-based model provides accuracy
of 78% in estimating the energy cost of the residential buildings. A method which
integrates BIM with ANN model for limiting the deviation between predicted and
Vietnam J. Comp. Sci. 2024.11:1-22. Downloaded from worldscientific.com

actual energy consumption rates is proposed in Ref. 45. This deviation is due to the
fact that energy performance tools, which are used to simulate the expected energy
consumption of a given building do not take into account the occupants' behavior.
In this study, optimum occupant behavior pro¯les that are generated via the ANNs
are inputted into the BIM ¯le in order to run necessary energy simulations to vali-
date the estimates of energy use of the building system. The results show that the
heating and cooling energy consumption can be predicted with an average of 10.4%
deviation from the actual performance measured in the buildings.
Automatically detecting clashes in designs by merging di®erent BIM models
supports the building design stage. In Ref. 46 six kinds of supervised ML algorithms
(decision tree, random forest, JRip-based rule methods, binary logistic regression,
naïve Bayes, and Bayesian network) are used to automatically distinguish relevant
and irrelevant clashes. Reference 47 develops a method that automatically ¯lters out
irrelevant clashes from the BIM-generated clash detection reports by combining the
two techniques of rule-based reasoning and supervised ML. BIM software is used for
clash detection and rule-based reasoning is used to preliminarily determine the type
of clashes. The results of this classi¯cation are added into the dataset of ML for
training and testing classi¯ers (k-NN, SVM, DT, random forest, bagging). The av-
erage predictive performance accuracy obtained by using the hybrid method is up to
96%. In Ref. 48 an AI system for design clash resolution was developed with ML and
heuristic optimizing techniques. A parallel backpropagation neural network was
employed as a learning procedure to extract the clash resolution knowledge patterns.
Based on these patterns, simulated annealing was adopted to develop a system for
automatically resolving design clashes from the constructor's perspective. To achieve
this in BIM models, the proposed system utilizes the API provided by the BIM
software as the portal to detect clashes, retrieve the geometric and spatial pro¯les of
elements, and revise the elements.
In Ref. 49 a natural language processing and deep learning-based approach for
supporting a BIM-based automated rule checking system is described. The semantic
analysis for building regulations is performed in two steps. First, the meaning of
words and the topic of sentences are learned by the neural net-based word embedding
 14

G. Slusarczyk & B. Strug

technique (word2vec model) and then sentence classi¯cation is conducted with a

deep learning model.

3.4. Building condition diagnosis

Semantically enriched BIM models are often used in the heritage industry to manage,
analyze and diagnose the assets at varying stages of the conservation process.
by 58.84.143.183 on 06/01/24. Re-use and distribution is strictly not permitted, except for Open Access articles.

Authors of Ref. 50 used ML techniques to automatically classify heritage buildings.

SVM are proposed to extract the main structural components such as °oors, ceilings,
roofs, walls and beams. The proposed semantic labeling of the objects is based on
features, which encode both local (surface area, orientation, dimensions) and con-
textual information (normal similarity, coplanarity, paralellity, proximity, topology)
Vietnam J. Comp. Sci. 2024.11:1-22. Downloaded from worldscientific.com

extracted form training datasets. The proposed automated feature extraction algo-
rithm combined with an SVM classi¯er takes the preprocessed data in the form of
planar triangular meshes and outputs the classi¯ed objects. The algorithm was
trained and tested using real data of a variety of existing buildings, like houses,
o±ces, industrial buildings, and churches. Ten structures representing di®erent types
of buildings were selected for the evaluation. The average accuracy of the model is
81%. The experiments prove that the approach reliably labels entire point cloud
datasets and can e®ectively support experts in documenting and processing heritage
assets.
ML methods are also used for defecting classi¯cation in masonry walls of historic
buildings.51 First, the process of Scan-to-BIM, which automatically segments point
clouds of ashlar masonry walls into their constitutive elements, is presented. Then
the ML-based approach to classi¯cation of common types of wall defects, which
considers both the geometry and color information of the acquired point clouds, is
described. The found defects are recorded in a structured manner within the BIM
model, which allows for monitoring the e®ects of deterioration. A supervised logistic
regression algorithm has been employed to classify di®erent types of decay using
parameters of roughness of stones and dispersion of color in stones. Stones labeled as
`defective' by experts are used for training the classi¯er, which is subsequently
employed to label new data. The proposed approach has been tested on data from the
main façade of the Royal Chapel in Stirling Castle, Scotland. For the training process
samples of three classes of decay (erosion, mechanical damage and discoloration)
were used. Fifteen samples (5 of each class) were included in the test set, obtaining a
global accuracy of 93.3% in the classi¯cation.

3.5. BIM enhancement in the FM context

Semantic enrichment by integrating FM information with a Building Information
Model is presented in Ref. 52. At ¯rst, various ML algorithms, which analyze the
unstructured text from occupant-generated work orders and classify it by category
and subcategory with high accuracy, have been investigated. Then, three learning
methods, TF, Term Frequency-Inverse Document Frequency (TF-IDF), and Random
 ML Methods in BIM-Based Applications 15

Forest classi¯er, were applied to perform this classi¯cation. A set of 155.00 historical
WOs was used for model development and testing textual classi¯cation. Classi¯er
prediction accuracies ranged from 46.6% to 81.3% for classi¯cation by detailed
subcategory. It increased to 68% for simple TF and to 90% for Random Forest when
the dataset included only the ten most common subcategories. An FM-BIM inte-
gration provides FM teams with spatio-temporal visualization of the work order
categories across a series of buildings and helps prioritize maintenance tasks. The
by 58.84.143.183 on 06/01/24. Re-use and distribution is strictly not permitted, except for Open Access articles.

paper shows that ML can be applied to support FM activities and can enhance BIM
use in the FM context.
To provide a good maintenance strategy for building facilities Cheng et al.
developed a data-driven predictive maintenance planning framework based on BIM
Vietnam J. Comp. Sci. 2024.11:1-22. Downloaded from worldscientific.com

and Internet-of-Things (IoT) technologies.53 Data collection and data integration

among the BIM models, FM system, and IoT network are undertaken in the infor-
mation layer, while ANN and SVM are used in the application layer to predict the
future condition of mechanical, electrical, and plumbing (MEP) components. BIM
models provide accurate and complete information for predictive maintenance, in-
cluding component type, dimensions, materials, capacity, location, and installation
year. An IoT sensor network was established to gather sensor data from building
facilities and environment during the operation period. The prediction models are
trained based on the continuously updated real-time sensor data and the accumu-
lated maintenance record. The input datasets are randomly divided into three sets:
80% for model training, 10% for cross-validation, and 10% for model testing. The
accuracy of prediction of MEP components condition is equal to 96.547% for SVM
and 96.422% for ANN. The results show that the constantly updated data obtained
from the information layer together with the ML algorithms in the application layer
can e±ciently predict the future condition of MEP components, prevent failure and
extend the lifetime of these components. In Ref. 54 a decision support system for
assessing the quality of a large building stock is proposed. The system is meant to
help owners of large building stocks in managing a huge number and variety of
buildings, taking into consideration a lot of aspects related to di®erent technical
issues, such as accessibility, energy e±ciency, life safety, ¯re protection, and seismic
vulnerability. It can extract relevant information from a BIM database and evaluate
the compliance of existing buildings with a set of rules by means of the application of
Bayesian networks.
A knowledge-based intelligent Building Energy Management System is consid-
ered in Ref. 55. To support the integration of multiple data sources and improve
interoperability, the solution makes use of BIM principles and semantic web tech-
nologies in the form of a holistic knowledge base, into which the rules are integrated.
The developed ontology is based on the IFC. The populated rules are a mix of an
intelligent rule generation approach using ANN and Genetic Algorithms, and also
data mining rules using decision tree techniques on historical data. The rules are
queried by the fuzzy reasoner together with the currently monitored state of the
building to generate suggestions for the FM.
 16

G. Slusarczyk & B. Strug

3.6. Knowledge extraction from BIM

Methods of extracting knowledge from existing BIM models with the use of ML are
also considered. In Ref. 2 the notion of the building ¯ngerprint is used to capture the
main characteristics of a building design. It serves as a measure of similarity which
allows for ¯nding a suitable reference for a given problem. The ¯ngerprint is based on
accessibility and adjacency relationships among spaces within a building model.
by 58.84.143.183 on 06/01/24. Re-use and distribution is strictly not permitted, except for Open Access articles.

Therefore, the authors retrieve accessibility and adjacency relationships among

spaces encoded in IFC models, and build accessibility and adjacency graphs.
Building ¯ngerprints are automatically generated based on a spatial semantic query
language for BIM and applied as indexes of the building model repository.
In Ref. 56 the ensembles of ANNs are used for cost estimation based on building
Vietnam J. Comp. Sci. 2024.11:1-22. Downloaded from worldscientific.com

information models. Building Information Model serves as a repository of data.

A regression model for cost estimation of building °oor structural frames, as higher-
level elements, was created. This approach was continued in Ref. 57, where a model
based on the support vector regression and radial basis kernel functions was devel-
oped and proposed to predict the construction costs of building °oor structural
frames. The ability to extract certain information about the building and structural
members of the °oor frames from BIM was combined with the capabilities of ML.
The training data for ML included inputs that represented features of the building
and structural members' belonging to the °oor structural frames and outputs that
represented corresponding real-life cost estimates of the °oor structural frames. The
obtained results show that the proposed model allows predicting costs with satis-
factory accuracy.
A BIM and ML integration framework for automated property valuation was
proposed.58 It contains a fundamental database interpretation, an IFC-based
information extraction and an automated valuation model based on genetic
algorithm optimized ML (the gradient boosting regression (GA-GBR)). The
value-relevant attributes or properties for property valuation were achieved
through quantitative analysis of 174 archival documents. The required value-speci¯c
design information was extracted automatically from an IFC-based BIM instance
model to support property valuation. The GA-GBR model was used to evaluate
property data in the American real estate market. The experimental dataset was
divided into 20 groups with 1000 houses data in each of them and randomly split into
training set (70%) and testing set (30%). The 56 input features were further used
for problem encoding in the GA-GBR training section. The results indicate that
GA-GBR model achieves an accuracy of 89.4%.
In Ref. 59, a Bayesian network is used to gain knowledge about existing bridges by
the use of data in the bridge management system in order to support bridge engi-
neering design. Two variants for the generation of the Bayesian network were con-
sidered: The Tree Augmented Naive Bayesian Network Algorithm and the manual
creation of the network by a knowledge engineer. Both variants permit the deter-
mination of design bridge parameters on the basis of given boundary conditions.
 ML Methods in BIM-Based Applications 17

The ¯rst variant ¯ts better to the generated training datasets, while the manually
created network supports the bridge design process in a more intuitive way. The
extracted knowledge is used for new bridge design problems to generate BIM models
for possible bridge variants.

4. Discussion
by 58.84.143.183 on 06/01/24. Re-use and distribution is strictly not permitted, except for Open Access articles.

In this paper, a signi¯cant number of recent papers related to the application of

di®erent AI methods in BIM have been presented. The papers have been carefully
selected by the authors rather than obtaining them via a keyword search of the
present state-of-the art results. There are several important observations as well as
Vietnam J. Comp. Sci. 2024.11:1-22. Downloaded from worldscientific.com

lessons that can be drawn from this analysis. The most important one is that the
majority of the best results are obtained not by using one method but rather by
combining di®erent approaches. The methods most often used are summarized in
Table 1. It can be observed that di®erent variants of neural networks and convo-
lution neural networks are the method of choice for the majority of the researchers.
Nevertheless, other methods like Bayesian classi¯er, decision trees, or genetic algo-
rithms were also used with success either on their own or, more often, in combination
with other methods.

Table 1. Methods of AI used for di®erent domains of applications in BIM.

Method Domain of application Papers

Neural networks Semantic classi¯cation 4,9,10,11

Supporting design 39,40,41,42,43,44,45,48
FM 53,55
Knowledge extraction 56
Convolution neural networks Semantic classi¯cation 14,19,23,26,28
Semantic enrichment 35,
Graph neural networks Semantic classi¯cation 27
Decision trees/forests Semantic classi¯cation 11
Semantic enrichment 36
Supporting design 46,47
FM 52
SVM Semantic classi¯cation 14,18
Semantic enrichment 33
Supporting design 38
Building condition 50
FM 53
Knowledge extraction 57
DBFs 3D model classi¯cation 17
Bayesian networks Supporting design 46
FM 54
Knowledge extraction 59
Genetic algorithms Supporting design 38
FM 55
Knowledge extraction 58
Rule-based methods Supporting design 46,47
 18

G. Slusarczyk & B. Strug

5. Conclusions
Traditional methods for modeling and optimizing complex structure systems require
huge amounts of computing resources. Arti¯cial-intelligence-based methods can
often provide valuable alternatives for e±ciently solving problems in architectural
and engineering design, construction and manufacturing.
ML techniques have considerable potential in the development of BIM. The
by 58.84.143.183 on 06/01/24. Re-use and distribution is strictly not permitted, except for Open Access articles.

application of classi¯cation algorithms would enable machines to do tasks usually

done by hand. Results from ML on architectural datasets provide a relevant
alternative view to explicit querying mechanisms and provide useful insights for
more informed decisions in the design and management of buildings. ML might
facilitate less experienced users to query complex BIM datasets for project-speci¯c
Vietnam J. Comp. Sci. 2024.11:1-22. Downloaded from worldscientific.com

insights.
It was shown that ML algorithms can learn the key features of a building
belonging to a certain category, and this acquired knowledge could be used in the
future when designing the methods to automatically design other structures based on
BIM historical data. The described techniques can be applied for retrieval, reference,
evaluation of designing, as well as generative design. The presence of historical data
combined with the acquired knowledge of the building type key features could help in
developing methods for automatically designing building structures with required
characteristics. The presented methods can be extended to further subdivide the
BIM main categories into sub-categories that could represent di®erent areas of in-
terest in these structures.
Many problems in architectural and engineering design, construction manage-
ment, and program decision-making, are in°uenced by many uncertainties, and
incomplete and imprecise knowledge. It seems that ML techniques should be able to
¯ll knowledge gaps in the knowledge bases and therefore they have broad
application prospects in the practice of design, construction, manufacturing, and
management. Object classi¯cation integrity is a fundamental requirement that
needs to be satis¯ed in order to ensure the e®ectiveness of using BIM applications.
Successful semantic enrichment tools would infer any missing information required
by the receiving application, thus alleviating the need for the domain expert to
preprocess the building model. They can help inexperienced users to solve complex
problems, and can also help experienced users to improve their work e±ciency and
share experience.
The authors believe that the widespread use of ML methods in BIM community
would require a more probabilistic and less deterministic approach to the parameters
in the models, and/or the implementation of correction and post-process measures.
Such measures could include automatic revision based on criteria established using
traditional programming, to pinpoint the elements where ML algorithms make
mistakes. In order to obtain consistent results with su±cient predictive potential, not
only is the choice of the right ML algorithms important, but also the choice, quan-
tity, and quality of the data used to train the algorithms.
 ML Methods in BIM-Based Applications 19

ORCID
Grażyna Ślusarczyk https://orcid.org/0000-0003-1032-1644
Barbara Strug https://orcid.org/0000-0002-2204-507X

References
1. W. Mazairac and J. Beetz, BIMQL? An open query language for building information
by 58.84.143.183 on 06/01/24. Re-use and distribution is strictly not permitted, except for Open Access articles.

models, Adv. Eng. Inform. 27 (2013) 444–456.

2. S. Daum, A. Bormann, C. Langenham and F. Petzold, Automated generation of building
¯ngerprints using a spatio-semantic query language for building information models, in
eWork and eBusiness in Architecture, Engineering and Construction: ECPPM 2014
(Vienna, Austria, 2014).
Vietnam J. Comp. Sci. 2024.11:1-22. Downloaded from worldscientific.com

3. C. Zhang, J. Beetz and M. Weise, Interoperable validation for IFC building models using
open standards, in Special Issue ECPPM 2014 10th European Conf. Product and
Process Modelling, Volume 20 (Vienna, Austria, 2014), pp. 24–39.
4. P. E. Núnez-Calzado, I. J. Alarc
on-L
opez and D. C. Martínez-G
omez, Machine learning
in BIM, in EUBIM 2018 (Valencia, Spain, 2018).
5. G. Ślusarczyk and B. Strug, Machine learning methods for BIM data, ACIIDS (2) (2022)
230–240.
6. BuildingSMART, (2013), http://www.buildingsmart-tech.org/.
7. T. Liebich, IFC 2x Edition 3 Model Implementation Guide (2009), https://standards.
buildingsmart. org/, 2009.
8. S. Wang, G. Wainer, V. Rajus and R. Woodbury, Occupancy analysis using building
information modeling and cell-devs simulation, DEVS13: Proceedings of the Symposium
on the Theory and Simulation, Vol. 26 (2013), pp. 1–8.
9. T. Bloch and R. Sacks, Comparing machine learning and rule-based inferencing for
semantic enrichment of BIM models, Autom. Constr. 91 (2018) 256–272, doi: 10.1016/j.
autcon.2018.03.018.
10. Ch. Jin, M. Xu, L. Lin and X. Zhou, Exploring BIM data by graph-based unsupervised
learning, 7th International Conference on Pattern Recognition app, i action and Methods
(Madera, Portugal, 2018), pp. 582–589.
11. Th. Krijnen and M. Tamke, Assessing implicit knowledge in BIM models with machine
learning (Design Modelling Symposium, Copenhagen, 2015) 397–406.
12. J. Beetz, I. Blmel, S. Dietze, B. Fetahui, U. Gadiraju, M. Hecher, T. Krijnen, M. Lindlar,
M. Tamke, R. Wessel and R. Yu, Duraark: Enrichment and preservation of architectural
knowledge, in How to Manage Data and Knowledge Related to Interpretative Digital 3D
Reconstructions of Cultural Heritage?, LNCS 10025 (Springer, 2015), pp. 231–255.
13. C. Emunds, N. Pauen, V. Richter, J. Frisch and C. van Treeck, IFCNet: A benchmark
dataset for IFC entity classi¯cation, in Proc. EG-ICE 2021, Universitaetsverlag der TU
Berlin, Berlin (2021), pp. 166–175 (in German).
14. F. Lomio, R. Farinha, M. Laasonen and H. Huttunen, Classi¯cation of building infor-
mation model (BIM) structures with deep learning, arXiv:1808.00601 [cs.CV].
15. A. Howard, M. Zhu, B. Chen, D. Kalenichenko, W. Wang, T. Weyand, M. Andreetto and
H. Adam. Mobilenets: E±cient convolutional neural networks for mobile vision appli-
cations, (2017).
16. K. He, X. Zhang, S. Ren and J. Sun. Deep residual learning for image recognition, in 2016
IEEE Conf. Computer Vision and Pattern Recognition (CVPR) (2016), pp. 770–778.
17. L. Wang, Z. K. Zhao and N. Xu, Deep belief network based 3D models classi¯cation in
building information modeling, Int. J. Online Eng. 11(5) (2015) 57–63.
 20

G. Slusarczyk & B. Strug

18. B. Koo, S. La, N. W. Cho and Y. Yu, Using support vector machines to classify building
elements for checking the semantic integrity of building information models, Autom.
Constr. 98 (2019) 183–194.
19. J. Kim, J. Song and J. K. Lee, Recognizing and classifying unknown object in BIM using
2D CNN, in Int. Conf. Computer-Aided Architectural Design Futures (Springer, Daejeon,
Rep. Of Korea, Springer Singapore, 2019), pp. 47–57.
20. C. Szegedy et al., Rethinking the inception architecture for computer vision, in Proc.
IEEE Conf. Computer Vision and Pattern Recognition (Las Vegas, USA, 2016),
by 58.84.143.183 on 06/01/24. Re-use and distribution is strictly not permitted, except for Open Access articles.

pp. 2818–2826.
21. J. Deng et al., Imagenet: A large-scale hierarchical image database, in Proc. Computer
Vision and Pattern Recognition (Miami, USA, 2009), pp. 248–255.
22. M. Abadi et al., Tensor°ow: Large-scale machine learning on heterogeneous distributed
systems, arXiv:1603.04467.
Vietnam J. Comp. Sci. 2024.11:1-22. Downloaded from worldscientific.com

23. B. Koo, R. Jung and Y. Yu, Automatic classi¯cation of wall and door BIM element
subtypes using 3D geometric deep neural networks, Adv. Eng. Inform. 47 (2021) 101200.
24. H. Su, S. Maji, E. Kalogerakis and E. Learned-Miller, Multi-view convolutional neural
networks for 3D shape recognition, in Int. Conf. Computer Vision Santiago, Chile, 2015,
doi: 10.1109/ICCV.2015.114.
25. C. Qi, H. Su, K. Mo and L. J. Guibas, Pointnet: Deep learning on point sets for 3D
classi¯cation and segmentation, Comput. Vis. Pattern Recognit. 1(2) (2017) 652–660,
doi: 10.1109/cvpr.2017.16.
26. B. Koo, R. Jung, Y. Yu and I. Kim, A geometric deep learning approach for checking
element-to-entity mappings in infrastructure building information models, J. Comput.
Des. Eng. 8(1) (2021) 239–250.
27. F. C. Collins, A. Braun, M. Ringsquandl, D. M. Hall and A. Borrmann, Assessing IFC
classes with means of geometric deep learning on di®erent graph encodings, in Proc. 2021
European Conf. Computing in Construction (2021).
28. H. Luo, G. Gao, H. Huang, Z. Ke, C. Peng and M. Gu, A geometric-relational deep
learning framework for BIM object classi¯cation, arXiv:2212.00942.
29. Y. Wang, Y. Sun, Z. Liu, S. E. Sarma, M. M. Bronstein and J. M. Solomon, Dynamic
graph cnn for learning on point clouds, ACM Trans. Graph. 38(5) (2019) 1–12.
30. A. Dosovitskiy et al., An image is worth 16  16 words: Transformers for image recog-
nition at scale, arXiv:2010.11929.
31. R. Sacks, L. Ma, R. Yosef, A. Borrmann, S. Daum and U. Kattel, Semantic enrichment for
building information modeling: Procedure for compiling inference rules and operators for
complex geometry, J. Comput. Civ. Eng. 31(6) (2017).
32. L. Ma, R. Sacks, U. Kattel and T. Bloch, 3D object classi¯cation using geometric features
and pairwise relationships: 3D object classi¯cation using features and relationships,
Comput.r-Aided Civ. Infrastruct. Eng. 33(2) (2018) 152164.
33. A. Xiong, B. Adan and D. H. Akinci, Automatic creation of semantically rich 3D building
models from laser scanner data, Autom. Constr. 31 (2013) 325–333.
34. D. Wolpert, Stacked generalization, Neural Netw. 5 (1992) 241–259.
35. C. Emunds, N. Pauen, V. Richter, J. Frisch and C. van Treeck, Sparse-BIM: Classi¯-
cation of IFC-based geometry via sparse convolutional neural networks, Adv. Eng.
Inform. 53 (2022) 101641.
36. R. Sacks, T. Bloch, M. Katz and R. Yosef, Automating design review with arti¯cial
intelligence and BIM: State of the art and research framework, in ASCE Int. Conf.
Computing in Civil Engineering 2019 (Atlanta, USA, 2019), pp. 353–360.
37. S. B. Kotsiantis, I. Zaharakis and P. Pintelas, Supervised machine learning: A review of
classi¯cation techniques, Emerg. Artif. Intell. Appl. Comput. Eng. 160 (2007) 3–24.
 ML Methods in BIM-Based Applications 21

38. B. Chen, Q. Liu, H. Chen, L. Wang, T. Deng, L. Zhang and X. Wu, Multiobjective
optimization of building energy consumption based on BIM-DB and LSSVM-NSGA-II,
J. Clean. Prod. 294 (2021) 126153, doi: 10.1016/j.jclepro.2021.126153.
39. S. Yuan, Z.-Z. Hu, J.-R. Lin and Y.-Y. Zhang, A framework for the automatic integration and
diagnosis of building energy consumption data, Sensors (Switzerland) 21(4) (2021) 1–19.
40. P. Geyer and S. Singaravel, Component-based machine learning for performance pre-
diction in building design, Appl. Energy 228 (2018) 1439–1453.
41. P. Geyer, M. M. Singh and S. Singaravel, Component-based Machine Learning for Energy
by 58.84.143.183 on 06/01/24. Re-use and distribution is strictly not permitted, except for Open Access articles.

Performance Prediction by MultiLOD Models in the Early Phases of Building Design,

Lecture Notes in Computer Science (including subseries Lecture Notes in Arti¯cial
Intelligence and Lecture Notes in Bioinformatics), Vol. 10863 (Springer Cham, 2018),
pp. 516–534.
42. G. Ma, Y. Liu and S. Shang, A building information model (BIM) and arti¯cial neural
Vietnam J. Comp. Sci. 2024.11:1-22. Downloaded from worldscientific.com

network (ANN) based system for personal thermal comfort evaluation and energy e±-
cient design of interior space, Sustainability (Switzerland) 11 (2019) 4972.
43. V. Gan, H. Luo, Y. Tan, M. Deng and H. Kwok, BIM and data-driven predictive analysis
of optimum thermal comfort for indoor environment, Sensors 21(13) (2021).
44. A. Alshibani and O. S. Alshamrani, ANN/BIM-based model for predicting the energy
cost of residential buildings in Saudi Arabia, J. Taibah Univ. Sci. 11(6) (2017) 1317–1329.
45. A. W. Hammad, Minimising the deviation between predicted and actual building per-
formance via use of neural networks and BIM, Buildings 9(5) (2019) 131.
46. Y. Hu and D. Castro-Lacouture, Clash relevance prediction based on machine learning, J.
Comput. Civ. Eng. 33(2) (2019) 04018060, doi: 10.1061/(ASCE)CP.1943-5487.0000810.
47. W. Y. Lin and Y.-H. Huang, Filtering of irrelevant clashes detected by BIM software
using a hybrid method of rule-based reasoning and supervised machine learning, Appl.
Sci. (Switzerland) 9(24) (2019) 5324, doi: 10.3390/app9245324.
48. H. C. Hsu, S. Chang, C. C. Chen and I. C. Wu, Knowledge-based system for resolving
design clashes in building information models, Autom. Construct. 110 (2020) 103001.
49. J. Song, J. Kim and J. K. Lee, NLP and deep learning-based analysis of building reg-
ulations to support automated rule checking system, Int. Association for Automation and
Robotics in Construction (I.A.A.R.C) (Berlin, Germany, 2018), pp. 586–592.
50. M. Bassier, M. Vergauwen and B. van Genechten, Automated classi¯cation of heritage
buildings for as-built BIM using machine learning techniques, ISPRS Ann. Photogramm.
Remote Sens. Spat. Inf. Sci., Vol. IV-2/W2 (2017), pp. 25–30.
51. E. Valero, A. Forster, F. Bosch
e, C. Renier, E. Hyslop and L. Wilson, High level-of-detail
BIM and machine learning for automated masonry wall defect surveying, in Proc. 35th
Int. Symp. Automation and Robotics in Construction (ISARC 2018) (2018).
52. J. McArthur, N. Shahbazi, R. Fok, Ch. Raghubar, B. Bortoluzzi and A. An, Machine
learning and BIM visualization for maintenance issue classi¯cation and enhanced data
collection, Adv. Eng. Inform. 38 (2018) 101–112.
53. J. C. P. Cheng, W. Chen, K. Chen and Q. Wang, Data-driven predictive maintenance
planning framework for MEP components based on BIM and IoT using machine learning
algorithms, Autom. Constr. 112 (2020) 103087, doi: 10.1016/j.autcon.2020.103087.
54. A. Corneli, S. Meschini, V. Villa, G. M. Di Giuda and A. Carbonari, A decision support
system for the multicriteria analysis of existing stock, 196 (2017) 682–689.
55. K. McGlinn, B. Yuce, H. Wicaksono, S. Howell and Y. Rezgui, Usability evaluation of a
web-based tool for supporting holistic building energy management, Autom. Constr. 84
(2017) 154–165.
56. M. Juszczyk, Implementation of the ANNs ensembles in macro-BIM cost estimates of
buildings' °oor structural frames, AIP Conf. Proc. 1946 (2018).
 22

G. Slusarczyk & B. Strug

57. M. Juszczyk, Cost estimates of buildings' °oor structural frames with the use of support
vector regression, IOP Conf. Ser.: Earth Environ. Sci. 222(1) (2019).
58. T. Su, H. Li and Y. An, A BIM and machine learning integration framework for auto-
mated property valuation, J. Build. Eng. 44 (2021).
59. D. Singer, M. Bugler and A. Borrmann, Knowledge based bridge engineering arti¯cial
intelligence meets building information modeling (2016).
by 58.84.143.183 on 06/01/24. Re-use and distribution is strictly not permitted, except for Open Access articles.
Vietnam J. Comp. Sci. 2024.11:1-22. Downloaded from worldscientific.com