Automatic generation of Structural models from BIM models using

Semantics and Machine Learning

Lini Xiang1, Gang Li1, Haijiang Li2

1
Dalian University of Technology, China, 2Cardiff University, UK
[email protected], [email protected], [email protected]

Abstract. Simulation of a BIM model under different disasters can provide reliable data for
designing safer structures, and precondition is to generate the matched finite element models from
the BIM model correctly. However, traditionally, obtaining the structural model mainly relies on
expert experience and manual programming, and once the scenario changes, the code needs to be
developed again. This paper introduces an automatic approach of mapping from BIM model context
to finite element parameters utilizing semantic embeddings and artificial neural networks. It can
generate a structural model immediately usable from BIM model via pre-trained models and
background knowledge, without manual adjustment. A reinforced concrete frame is used to verify
the validity of the proposed method. The results show that new tools of artificial intelligence are
able to solve the challenges of applying BIM data in expert models, but it needs more evidence in
improving efficiency, coding workload, automation, and generalization capability.

1. Introduction
The finite element (FE)-based uncertainty quantification is widely used in many
computationally intensive tasks, such as structure design in civil, mechanical, and aerospace
fields (Melchers et al., 2017). However, the application of BIM models for uncertainty analysis
and structural reliability is hindered due to the difficulty of generating structural models from
BIM models. Compared with carbon neutrality and structural health monitoring, less attention
and research have been devoted to this area.
There have been many attempts to convert BIM models to finite element models. The technical
challenges they faced were how to extract information from Industry Foundation Class (IFC)
files to compose simulation-oriented models - from a uniform information exchange format
across the domain to a proprietary format for a single domain; from serialized information
(Step), including component types, geometric 3D data, and physical properties, to an analysis
language (MATLAB, python); from descriptions based on real objects to descriptions based on
mathematical abstractions. The spread of BIM models around the world has increasingly
revealed the lack of means for structural engineers to make efficient use of BIM models.
Machine learning (ML) is a subfield of artificial intelligence, where algorithms are developed
to learn from data, generalize and make predictions (Mitchell, 1997). Under the right
development, ML models are able to bypass the difficulty of directly converting IFC instances
into structural model elements, as we use statistical learning instead of descriptive semantic
mapping to construct geometry, sectional parameters, and material parameters. This process
does not require human intervention or compatibility with different structure interpreters. In
addition, semantic web, ontologies and knowledge graphs are validated in capturing and reusing
domain knowledge from BIM models (Chen et al., 2022, Johansen et al., 2022,
Shahinmoghadam et al., 2022). Therefore, the BIM model is the natural data source for ML
training, providing semantic texts on building components as well as parameters.
This paper combines the advantages of both approaches and proposes a method to perform
structural element mapping according to the context of BIM models, enabling the generation of
1
customized structural models in shorter time and with less hard-coding. In Section II, introduce
how we prepare BIM training data, build artificial neural networks (ANNs), perform mapping
of components to FE elements and assemble them into OpenSees models. In Section III, a
reinforced concrete (RC) frame is used to validate the proposed method. Sections IV and V
summarize the key findings of this study and implications for similar work.

2. Related Work

There are three different ideas to solve the challenge of how to convert serialized building
information in IFC files to FE models. First, shape inference. The main work is to create
complex geometry and mesh it by surface reconstruction algorithms from 3D polylines and 3D
pointcloud data (Xu et al., 2019, Park et al., 2020, Rasoulzadeh et al., 2022). The reconstructed
geometries with volume mesh are then paired with boundary conditions, load and material
information to compose a structural analysis model. This field focuses on 3D modeling,
efficient algorithms for point-to-surface, and meshing, so it is suitable for dealing with large
structures with complex surfaces.

Second, rule-based reasoning. In the past time, many studies used semantic information from
BIM models as middleware to translate structural modeling intents (Ramaji et al., 2018, Sibenik
et al., 2020, Sibenik et al., 2021, Jia et al., 2022). Because serialized building information (Step
file) can provide a shared network of concepts and relationships, often presented in a graphical
form, this is the Semantic Web. Computers are able to reason about the Semantic Web to obtain
logical facts using the Web Ontology Language (OWL), which is called rule-based reasoning
process. An approach is to create IFC-to-structural mapping rules directly. Or use algorithms
to solve constraint graphs (a graph in which the parameters, shapes, and functions of the IFC
components form the graph elements while the constraint relationships form the network)
(Kirchner, 2022). By solving the (n+1)B-consistency problem, constraint graphs can be
transformed into a set of equations and variables of the structure. But these methods rely too
much on formalization knowledge and symbol reasoning, therefore it is the least efficient and
too abstract to understand.

Third, API interface development. The most direct and least compatible method is to use hard
coding to rewrite process of defining BIM model information to SAM files (Alirezaei et al.,
2016, Zhang et al., 2017), such as from Revit to CATIA. This approach has high accuracy, but
low efficiency and poor generalization capability to reuse between different structural analysis
software (e.g. ETABS, MIDAS, SAP2000).

Considering the drawback of the above methods, we would like to improve in terms of
efficiency, coding workload, automation, generalization capability, and ease of understanding.
There is a good prior study that is a set of machine learning-based BIM-to-OpenSees code
developed by the NHERI Center in the United States (Wang et al., 2019). However, they only
focused on machine learning and did not play the role of building semantics. This paper is
aimed to propose a new solution that takes advantage of both sides to achieve a better
conversion from BIM model to FE model.

2
3. Methodology

This section will introduce the types of structural analysis, the establishment of ML models,
and the construction of FE models. Figure 1 illustrates the complete flowchart.

Figure 1: The flowchart of generating FE models automatically from BIM models

3.1 Structural Analysis Concept

Structural Dynamics Equations. The following equation is a general form for calculating the
structural dynamic response (Vurtur Badarinath et al., 2021):
(t )  Cu (t )  Ku(t )  f
Mu (1)
where u(t ) is the displacement vector containing the nodal degrees of freedom depending on
. ..
time t , and u(t ) and u(t ) are its first and second derivatives of time, respectively representing
the velocity and acceleration. M is the mass matrix, C is the damping matrix, K is the
stiffness matrix, and f is the vector of discretized loads applied to each degree of freedom.
M , C , and K are determined subsequently after the geometric form and material properties
are determined. Given the loading vector f , time history analysis is conducted to obtain the
discrete response data of each degree of freedom at each time point:
.
. .
 ..  .. ..

f =  f 1 , ,f n   u =  u1 , ,un  ,u =  u1 , , un  ,u =  u1 , , u n  (2)
   
OpenSees FEA Model. In this paper, the RC building is simplified as a 2D shear model,
considering only the response along the lateral direction. This multi-story, multi-bay frame with
gravity system comprise nodes, elastic beam elements and nonlinear column elements. FE
analysis is conducted using the OpenSees fiber elements that are discretized into the
longitudinal reinforcing bars and the concrete fibers wrapped outside. Rigid beam-column
connections are adopted. The degrees of freedom of the bottom nodes are restricted to zero.
OpenSees variables are considered include:
 Geometrical parameters: the cross-sectional area of reinforcing bar (As), column
height (height), beam span (width), and thickness of concrete protective layer (cover).

3
  Material parameters: yield strength of the steel (fy), elastic modulus (E), strain
hardening coefficient (b) and compressive yield strength of concrete (fc).

3.2 Machine Learning Training

Data Preparation. The training process is to map the beam and column entities of the BIM
model directly to the necessary structural parameters. Considering that all entities and their
properties are usually exported in RDF or OWL format for querying and reasoning in a natural
language manner, there exists a gap making BIM entities not applicable to ML processes.
Fortunately, the BIM community has developed a method to generate embeddings of semantic
texts inside knowledge graphs in RDF or OWL format (Shahinmoghadam et al., 2022). The
embedding is a feature vector representation of semantic information to be used for digital tasks.
In order to convert beam and column instances in BIM model to ML training data, embedding
learning is conducted:
1. SPARQL query is used to extract all beam and column entities, as well as related
mechanical properties (listed in the above section) entities, in turtle serialization.
2. The pyRDF2Vec library (Vandewiele et al., 2022) is used to generate feature vectors
through *.ttl file, with 100-dimension by default for each entity.
3. t-SNE (t-distributed Stochastic Neighbor Embedding) algorithm is employed to reduce the
dimension of the generated embeddings. Since large number of irrelevant dimensions
usually cause overfitting in training, 2-dimension embeddings are used in our regression
tasks.

ANNs Model. ANNs are complex models that emulates the structure and function of neurons
in the human brain to solve real-world complex problems. It uses neurons and weights to
represent the connections of data, and uses activation functions to tune the output of neurons
(Kelleher et al., 2020). We employ a supervised learning on a specific dataset to generate the
expected output for a set of paired inputs, to build regression models based on the embeddings
generated from structural components and their mechanical properties in the BIM model, to
predict parameters of new beam/column instances. The learning process is accomplished by
adjusting the network weights and various hyperparameters (learning rate, hidden layer, batch
size) to minimize the observed errors (Kelleher et al., 2020). In this paper, it was trained with
5000 epochs using Adam optimizer, with Rectified Linear Unit (ReLu) activation function,
learning rate of 0.0001, and batch size of 32. The details of the ANNs are shown in Table 2. All
code was developed by Keras library in Python.

Table 1: Details for ANNs structure in this study.

Layer Description Parameter

Input Layer Factors Beam embedding / Column
embedding
Number of Units 1
Number of Hidden Layers 4
Hidden Layer Number of Units 512, 64, 64, 64
Activation Function ReLu
Output Layer Dependent Variables width,fy,E,b,As / height,fc,cover
Number of Units 5/3
Error Function Mean Squared Error

4
For regression tasks in this study, the quality of predictions is quantified by mean absolute error
1 n 1 1 n1
(MAE, MAE   yi  yˆ i ), mean squared error (MSE, MSE    yi  yˆ i  ), and R-
2

n i 0 n i 0
n

 y  yˆ i 
2
i
squared (R2, R 2  1  i 1
n
). ANNs with higher R2 closer to 1 and lower MAE and
 y  y
2
i
i 1
MSE closer to 0, are better fitted.

3.3 OpenSees Model Generation

Given the embeddings, and after the prediction via pre-trained ANNs, output features in one-
dimension are obtained. The ultimate goal is to use these features to auto-write a reality-
compliant FE model. A simple regression (ordinary least squares, OLS) is conducted to map
the features to the real-world parameters. The component parameters will be fed into the
template of the 2D shear frame model in the OpenSeesPY library. In addition to our method,
Tcl builder written in C++ language (Wang et al., 2019) is able to write FE model in Tcl
language.

4. Results

In this section for validation, a BIM model of a 10-story, 14-bay RC frame structure is used
(see Figure 2). The BIM model was serialized by IFCtoRDF tool (Pauwels et al., 2016, Pauwels,
2020) containing the basic building component types, geometric and spatial information,
physical properties, and various instances.

Two different datasets are created, corresponding to beam elements, beam parameters and
column elements, column parameters, respectively. Inside the datasets, the labels are
embeddings of extracted properties, while the data are the embeddings of the entities. Each
dataset is split into training samples (80%) and test samples (20%). The summary of the
building ontology and the two datasets are shown in Table 2.

Figure 2: BIM model of multi-story RC frame in Shenyang, China

5
 Table 2: Summary statistics of the building ontology and two datasets.

Description of the Description of the Dataset to be trained

Knowledge Graph
Triples Maximum Train/Test Data Label
LoD
samples
Beam 1360 7 Beam 35 44 [44×5]
ontology dataset (entity) (embedding width,
9 fy, E, b, As)
Column 64948 - Column 312 390 [390×3]
ontology dataset (entity) (embedding height,
78 fc, cover)

In the first step of the experiment, embeddings were calculated. Figure 3 shows the 2D
visualization of embeddings of beam/column entities and embeddings of width/height
parameters by t-SNE algorithm. It can be observed that there are spatial pattern similarities
between the beam and beam parameters. The scatter of column and column parameters,
however, shows a tendency to be uncorrelated. The embeddings of the other parameters are not
shown here.

Figure 3: 2D visualization of embeddings of (a) beam and its width (b) column and its height

The ANN training process is shown in Figure 4 and Figure 5. The left graph shows the gradual
convergence of MSE as the epoch increases, while the right graph shows the error between
predictions and true values of test samples. More performance metrics on ANNs are
summarized in Table 3. Regarding MAE and MSE, the values for beam-ANN model are 0.3288
and 0.1764, respectively, and for column-ANN model are 1.0188 and 1.4909. All the above
indicators are close to 0, indicating that hyperparameters are correctly chosen and there is no
overfitting or underfitting. However, the R2 for the beam-ANN model is 0.3389, indicating that
only 33% of the variation in variables were predicted. The R2 of the column-ANN model is
0.0047, indicating that the prediction accuracy of this ANN model is similar to that of taking
the sample mean directly. Considering both samples are small and dimensionality reduction
and the ANNs are the same, the difference in training results between the two datasets can only
be interpreted as the fundamental difference in the semantic text, i.e., whether there is an
explicit linking path (or implicit one-to-one relation) in the target ontology that can connect the
entities to the physical properties.

6
 Figure 4: The training loss curves and the prediction of Beam-width

Figure 5: The training loss curves and the prediction of Column-height

Table 3: Performance of the two ANNs on the testing set.

Metrics Beam-ANN model Column-ANN model

Mean Absolute Error 0.3822 1.0188
Mean Squared Error 0.1764 1.4909
R-squared 0.3389 0.0047

Inst:IfcBeam_449008 and inst:IfcColumn_141474 as entities were introduced into the pre-

trained ANNs, to verify how effective the semantic text is in communicating modelling intents.
Table 4 summarized input variables and output prediction. Comparing the predictions of both
sides, it is obvious that beam semantics-communicated parameters is learned better.

Table 4: Randomly input variables and the corresponding ANNs prediction.

Input Output variable

Entity variable/Embedding Width Fy E b As
(inch) (kips) (kips) (in
^2)
IfcBeam_4490 [1.18580759, Actual 149.61 60 30000 0.01 0.6
data

7
 08 3.09675908] Prediction 146.80 57.96 29537. 0.03 0.6
48 4
Height Fc Cover
(inch) (kips) (inch)
IfcColumn_141 [-2.41622810, Actual 165.35 5 0.98
data
474 1.19574952] Prediction 159.98 6.72 1.14

The obtained parameters were filled into OpenSeesPY code to perform reliability analysis
under ground motion. Given that time history analysis is time-consuming, the computational
cost of large-scale uncertainty simulation is unacceptable if performed with Monte Carlo
simulation. Here we chose 50 Latin hypercube sampling (LHS). For each LHS, all predicted
structural parameters were set as Gaussian variables, to perform time history analysis under the
"elCentro" acceleration record, to calculate the maximum displacement of the top floor. After
50 loops, see if any displacement record exceeded 1% of the total height. Figure 10 shows the
uncertainty propagation results. The simulation results show that all displacements are less than
16 inches (1% of the total height), but computing failure probability needs more simulations.
We only verified the success BIM-FEA conversion process, not accurate failure probability.

Figure 6: Earthquake anlysis for the generated FE model (Uncertainty propagation results for
the maximum displacement of the top layer)

5. Discussion
Section 4 showed the flowchart results step by step. The experiment demonstrated that, it is
feasible to achieve prediction of FE parameters and to automatically construct FE models
consistent with BIM models, by statistical learning based on the embeddings of building
ontology. The validity of the new method was verified. However, it should also be seen that the
prediction of the machine learning model in this paper should be further improved. Table 4
shows that the proposed ANN model has satisfactory accuracy in predicting the beam
parameters, although R2 is not close to 1 as desired (R2=0.3389). But the effect on column
entities (R2=0.0047) is only slightly better than directly using the average value to replace the
prediction process. Increasing the learning sample size may be necessary.

8
In addition, further comparisons of other methods mentioned in related work are needed to
verify the degree of improvement in terms of efficiency, coding workload, automation and
generalization capability. In particular, compared with the rule-based inference approaches,
how much work does our method decrease in getting specific properties of the specified
building entity. Besides, compared with pure machine learning, how better does learning and
prediction results be by using context data sources or building semantics. Finally, regarding
generalization capabilities, the proposed solution is expected to be adaptable when engineering
scenarios change, so more work is needed to demonstrate that it is not limited to being applied
to frame structures.

6. Conclusion
In this paper, we propose an automatic transformation approach from BIM models to structural
analysis models, using BIM ontology and artificial neural networks. The results demonstrate
that building semantics are able to convey implicit modeling knowledge and simulation intent
by machine learning process, and finally to construct the responding FE model.
The value of this study is that it provides a better solution to achieve the conversion of BIM
models to finite element models - using new tools of artificial intelligence to solve the
challenges of applying BIM data in expert models. And, compared with similar studies, there
is more space for improvement in our approach, as it relies more on machine intelligence rather
than human manual work. However, the drawback of this study is that only the effectiveness of
the new workflow has been validated, and more perceived potential advantages are not yet
supported by sufficient evidence. Therefore, future work includes completing comparisons with
similar studies and different types of conversion methods.

Acknowledgements
This research was funded by the National Natural Science Foundation of China (Grant No.:
11872142). The BIM model was provided by Associate Professor Haiyan Lu of Shenyang
University of Technology.

References
Alirezaei, M., Noori, M., Tatari, O., Mackie, K.R. and Elgamal, A. (2016). BIM-based Damage
Estimation of Buildings under Earthquake Loading Condition. Procedia Eng., 145, pp. 1051-1058.
Chen, X., Saluz, U., Staudt, J., Margesin, M., Lang, W., and Geyer, P. (2022). Integrated data-driven
and knowledge-based performance evaluation for machine assistance in building design decision
support. In: 29th International Workshop on Intelligent Computing in Engineering, EG-ICE 2022,
Aarhus, Denmark.
Jia, J., Gao, J., Wang, W., Ma, L., Li, J., and Zhang, Z. (2022). An Automatic Generation Method of
Finite Element Model Based on BIM and Ontology. Buildings, 12(11), 1949.
Johansen, K., Schultz, C. and Teizer, J. (2022). BIM-based Fall Hazard Ontology and Benchmark Model
for Comparison of Automated Prevention through Design Approaches in Construction Safety. In: 29th
International Workshop on Intelligent Computing in Engineering, EG-ICE 2022, Aarhus, Denmark.
408-417.
Kelleher, J.D., Namee, B.M. and D'arcy, A. (2020). Fundamentals of Machine Learning for Predictive
Data Analytics: Algorithms, Worked Examples, and Case Studies (2nd Edition), Cambridge, MA, The
MIT Press.

9
Kirchner, J. (2022). Computation of Ranges in Interval-based Constraint-Geometry of Building Models.
In: 29th International Workshop on Intelligent Computing in Engineering, EG-ICE 2022, Aarhus,
Denmark. 206-215.
Melchers, R.E. and Beck, A.T. (2017). Measures of Structural Reliability, Hoboken, NJ, John Wiley &
Sons.
Mitchell, T.M. (1997). Machine Learning, New York, McGraw-Hill.
Park, S.I., Lee, S.H., Almasi, A. and Song, J.H. (2020). Extended IFC-based strong form meshfree
collocation analysis of a bridge structure. Autom. Constr., 119, 103364.
Pauwels, P. (2020). IFCtoRDF (Version 0.4). https://github.com/pipauwel/IFCtoRDF, accessed April
2023.
Pauwels, P. and Terkaj, W. (2016). EXPRESS to OWL for construction industry: Towards a
recommendable and usable ifcOWL ontology. Autom. Constr., 63, pp. 100-133.
Ramaji, I. J. and Memari, A. M. (2018). Interpretation of structural analytical models from the
coordination view in building information models. Autom. Constr., 90, pp. 117-133.
Rasoulzadeh S., Senk V., Kovacic I., Reisinger J., Füssl J., Hensel M. (2022). Linking Early Design
Stages with Physical Simulations using Machine Learning Structural Analysis Feedback of
Architectural Design Sketches. In: 29th International Workshop on Intelligent Computing in
Engineering, EG-ICE 2022, Aarhus, Denmark. 216-226.
Shahinmoghadam, M., Motamedi, A. and Soltani, M. (2022). Leveraging Textual Information for
Knowledge Graph-oriented Machine Learning: A Case Study in the Construction Industry. In: 29th
International Workshop on Intelligent Computing in Engineering, EG-ICE 2022, Aarhus, Denmark.
259-268.
Sibenik, G. and Kovacic, I. (2020). Assessment of model-based data exchange between architectural
design and structural analysis. J. Build. Eng., 32, pp. 101589.
Sibenik, G., Kovacic, I., Petrinas, V. and Sprenger, W. (2021). Implementation of Open Data Exchange
between Architectural Design and Structural Analysis Models. Buildings, 11, pp. 605.
Vandewiele, G., Steenwinckel, B., Agozzino, T. and Ongenae, F. (2022). pyRDF2Vec: A Python
Implementation and Extension of RDF2Vec. arXiv. https://arxiv.org/abs/2205.02283, accessed April
2023.
Vurtur Badarinath, P., Chierichetti, M. and Davoudi Kakhki, F. (2021). A Machine Learning Approach
as a Surrogate for a Finite Element Analysis: Status of Research and Application to One Dimensional
Systems. Sensors, 21, pp. 1654.
Wang, C., Jiang, C.G., Yu, S.X., Mckenna, F. and Law, K.H. (2019). NHERI-SimCenter/BIM2SAM.AI:
Release v1.0 (Version 1.0). Zenodo. http://doi.org/10.5281/zenodo.3509957, accessed April 2023.
Xu, Z., Rao, Z., Gan, V.J.L., Ding, Y., Wan, C., and Liu, X. (2019). Developing an Extended IFC Data
Schema and Mesh Generation Framework for Finite Element Modeling. Adv. Civ. Eng., 2019, 1434093.
Zhang, X.Y. and Hu, Z.Z. (2017). Research on model conversion approach towards structural finite
element analysis. Gongcheng Lixue/Eng. Mech., 34, pp. 120-127.

10