See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/361632164

NTAB_0: Design Priors for AI-Augmented Generative Design of Network

Tied-Arch-Bridges

Conference Paper · September 2022

CITATION READS

1 41

5 authors, including:

Sophia Kuhn George Klonaris

ETH Zurich ETH Zurich
2 PUBLICATIONS   2 CITATIONS    2 PUBLICATIONS   158 CITATIONS   

SEE PROFILE SEE PROFILE

Rafael Bischof Walter Kaufmann

Swiss Data Science Center ETH Zurich
5 PUBLICATIONS   4 CITATIONS    108 PUBLICATIONS   1,497 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Rethinking Structural Concrete for Digital Fabrication View project

XR Methods for AEC View project

All content following this page was uploaded by Michael Kraus on 06 October 2022.

The user has requested enhancement of the downloaded file.

N T AB0 : Design Priors for AI-Augmented Generative Design of Network Tied-Arch-Bridges 1

N T AB0: Design Priors for AI-Augmented

Generative Design of Network Tied-Arch-Bridges
Sophia V. Kuhn1 , Rafael Bischof2 , Georgios Klonaris1 , Walter Kaufmann1 and Michael A. Kraus1
1 Chair of Concrete Structures and Bridge Design, ETH Zurich, Stefano-Franscini-Platz 5, 8093 Zurich,
Switzerland
2 Swiss Data Science Center, Turnerstrasse 1, 8092 Zürich, Switzerland

E-mail(s): {sophia.kuhn,kraus}@ibk.baug.ethz.ch

Abstract: Projects in the Architecture, Engineering and Construction (AEC) industry inherit a great
complexity due to a tremendous amount of design parameters, multiple objectives, and many involved
stakeholders. This research is concerned explicitly with the conceptual design stage of bridges,
where an in-detail analysis of many performance attributes (e.g. structural analysis) for each design
alternative is time-consuming and infeasible under the current approaches. In industry, the initial
design solution today is predominantly dependent on the prior knowledge and expertise of the involved
team. In contrast to the status quo, this paper introduces the novel concept of bridge design prior
models to predict the layout and structural properties of bridges as (near-optimal) starting point for
Generative Design (GD). The concept of design prior models for bridges is demonstrated on network
tied-arch bridges (NTAB), which can be described by a parameter set θB ∈ Rd . The prior N T AB0
takes a subset of design parameters θB1 ∈ Rd1 to generate meaningful suggestions for the remaining
parameters θB2 ∈ Rd−d1 at once or in sequence. N T AB0 is calibrated upon a curated database
consisting of existing real-world NTABs and captures numeric, semantic, and topological relations
between bridge properties such as materials, cross-sections or bracing systems. Calibration of
N T AB0 consists of two subsequent steps, where first clustering analysis is performed by applying the
k-Prototype as well as DBSCAN algorithms. In the second step, a predictive (discriminative) model
is trained using the gradient boosted decision tree algorithm from CAT BOOST. A subsequent study
evaluates the suitability of the algorithms to serve as sensible design priors. We found that the AI prior
model N T AB0 is able to suggest meaningful design parameters, assisting the designing team with an
informed initial bridge design for further design space exploration and optimisation. The application of
the AI prior model shows great potential to improve future construction projects by providing easy and
fast access to the information saved in the successfully built structures of today. Our approach brings
in additional expertise at an early design stage, enabling designers to make more informed decisions
towards optimised bridge structures.

Keywords: Conceptual Structural Design, Machine Learning, Design Automation, Algorithms

33. Forum Bauinformatik, München, 2022

2 Sophia V. Kuhn, Rafael Bischof, Georgios Klonaris, Walter Kaufmann, Michael A. Kraus

1 Introduction

In current Architecture, Engineering and Construction (AEC) practice, the conceptual design phase is
often disconnected from performance evaluations of the structure with respect to safety, functionality,
cost, material impacts, and the construction processes. However, this phase is most influential for a
building project. Today, the initial design is mainly based on the prior knowledge of (i) the designing
team, and (ii) the manual analysis of a few similar reference projects. This design practice results in
lengthy and costly processes with numerous manual translations, detailing, and reiterations of rigid
design models between early-stage designers and engineers or contractors of later planning stages.

Built bridge structures however represent the final result of detailed investigations during a bridge
project, in which the expertise of civil engineers was applied. This fact motivates the derivation of a
method for data collection and automated analysis as well as calibration of quantitative prior design
models to be used in conceptual design phases. Specifically, this paper proposes to (i) start free-
access and open databases on structural properties of bridges, and (ii) use modern Machine Learning
(ML) algorithms for data mining and calibration of predictive prior models upon these databases to
speed up early phase design and/or validate the design process choices. This idea is investigated
within the Bridge Genome Project, cf. Fig. 1, for different bridge types. The prior models and their
insights are interpreted as the "genome" of how bridges have been planned and built. The workflow
and results for calibrating a design prior model upon a repository are presented for the case of network
tied-arch bridges (NTAB), described by a parameter set θB ∈ Rd , leading to the N T AB0 prior.

Figure 1: The Bridge Genome Project Figure 2: Bridge properties

2 Background and Related Literature

Bayesian methods are an alternative approach to traditional statistical analysis. They are employed
widely for analysing and interpreting data or training Artificial Intelligence (AI) models in many fields,
including data and computer science as well as civil engineering [1]. The Bayesian method computes
a posterior probability about a problem based on observed data (likelihood) and prior knowledge using
Bayes’ theorem, cf. Sec. 3. Prior information is obtained from the results of previous experiments,
simulations, domain theory, and expert assessments [2]. Data-driven learning (i.e. learning from
data repositories or sets) through Machine and Deep Learning allows for new discriminative and

33. Forum Bauinformatik, München, 2022

N T AB0 : Design Priors for AI-Augmented Generative Design of Network Tied-Arch-Bridges 3

generative models to be employed to the design task. Discriminative models learn to solve a learning
task (e.g. classification, regression, clustering, dimensionality reduction) where for a new input the
respective output is provided. Applying a Bayes view, a discriminative model strictly learns the posterior
probabilities of the output given an input [2]. In our present work, the inferred discriminative models
serve as prior models for the design task. Here the posterior probabilities of the design output for a
given input (i.e. a specific bridge construction project situation) are augmented by the informative prior.

The potential of repositories, evolutionary algorithms, and parametric tools as drivers for design actions
was explored for performance-based shading device design for office buildings by [3]. [4] is one of the
few studies on bridge simulation data mining as an approach to extract knowledge and decision rules
from a database of pre-computed models (design variants) containing simulation results. However,
the greatest current hurdle for data mining employing AI in the AEC sector is the lack of systematized,
free-accessible databases to learn from the successes and failures of past engineering projects. The
potentials and pitfalls of data mining in the AEC sector are addressed in [5].

3 Methods

The formal capturing and quantitative incorporation of domain-knowledge in the form of priors distri-
butions of data about a design problem works within the Bayesian framework: P(D|M(θ)), where
P denotes a probability distribution, D is the data set and M(θ) the parametric model with design
parameters θ ∈ Rd respectively. Instead of employing Bayesian model averaging, we suggest to
a-priori select a model M∗ (θ), and subsequently the observed data set is a manifestation of that model
drawn according to the probability P(D|M∗ (θ)). The goal of the ML algorithms then is to estimate
P(M∗ (θ)). This paper applies the described approach to NTABs for calibrating the N T AB0 prior.

3.1 Database of Network Tied-Arch Bridges (NTABs)

In the AEC Industry, the systematic documentation of structural data (such as cross-sections, material
grades, bracing system types, etc.) of completed projects is currently still insufficient due to the
absence of public repositories. To nevertheless have a basis for creating a prior model for bridge
design, a data set was curated, which is publicly available via [6]. The database is structured in a table
containing mixed-data-type properties including functionality, geometrical, and material parameters of
203 NTABs. Fig. 2 shows the important bridge features together with a distinction of the categorical
resp. continuous data types. As the curated data set exhibits missing values, two pre-processing
methods are followed to gain a complete data set: for the first data set, the data is filtered for complete
rows (bridges). For the second, the missing values are filled utilizing a k-NN algorithm [7]. k-NN
estimates missing data by computing a weighted average over the k nearest neighbors in the data set
based on the euclidean distance between non-missing entries. We empirically found k = 3 to yield a
good compromise between consistency and diversity in the data set.

33. Forum Bauinformatik, München, 2022

4 Sophia V. Kuhn, Rafael Bischof, Georgios Klonaris, Walter Kaufmann, Michael A. Kraus

NTABs belong to the class of (tied-)arch bridges and have been first introduced by Per Tveit in the
1950s [8]. Arches are highly efficient structures due to their ability to carry loads by compression in
the case that the thrust line lies within its cross-section [9]. The horizontal forces carried by the arch
are short-circuited through the deck in tension. The structure typically includes two arch ribs to which
the vertical loads acting on the deck are suspended by inclined and intersecting hangers.

3.2 Machine Learning (ML) Algorithms

Our approach suggests to (i) identify structure (i.e. similarity groups) within the bridge data using
clustering algorithms, and (ii) train a discriminative ML model for predicting suitable bridge parameters.

The cluster analysis is applied to the filtered data set. As the NTAB data set carries mixed data
types, the k-Prototype algorithm as implemented in [10] is applied. The k-Prototype algorithm
combines the well-known k-means algorithm [11] with the k-modes algorithm [12], to handle mixed
data types of continuous and categorical data. To enable the identification of well separated clusters
in all dimensions the equal contribution of all properties to the clustering process is achieved by (i)
individually standardizing each property column to zero mean and unit variance, and (ii) by iteratively
choosing a relative importance weight γ so that an even contribution of both the continuous and
categorical properties is observed within the loss function of the k-prototype algorithm. Furthermore,
the clustering is conducted 100 times in parallel with a new ’Cao’ initialisation each time. The number
of clusters k is chosen according to the ’elbow method’. In order to test the stability of the clusters
induced via distance-metrics by the k-prototype algorithm, the Density-Based Spatial Clustering of
Applications with Noise (DBSCAN) [13] algorithm is used to provide a clustering of the data by fitting
probability distributions. However, the results are not shown within this paper.

The second step is to calibrate mixed-data-type ML algorithms upon the NTAB data set as a dis-
criminative design prior for predicting structural bridge properties for a specific project. This means,
that starting from a set of fixed design parameters θB1 ∈ Rd1 , the algorithm generates meaningful
suggestions for the remaining parameters θB2 ∈ Rd−d1 in a specific order, taking into account all the
previous predictions that have been made. For the data set at hand, a decision tree approach is
chosen as these algorithms are capable of solving both classification as well as regression tasks,
which is necessary due to the mixed-type nature of the data set. Specifically, we calibrate a CatBoost
[14] algorithm to the bridge data set as in a pre-study it outperformed other existing state-of-the-art
implementations of gradient boosted decision trees (such as XGBoost) in terms of accuracy. This is in
accordance with literature, where CatBoost was found especially efficient and accurate if categorical
features are present and play an important role [15]. The model is trained on 85% of the bridges from
the second imputed NTAB data set. The residual 15% of samples are used to test model performance.
A stratified train-test splitting method was applied. For the regression head, CatBoostRegressors
[14] are fitted to the training data using the Root Mean Squared Error (RMSE) loss function with an
evaluation via the RMSE on the test data set. For the classification head, CatBoostClassifiers [14]

33. Forum Bauinformatik, München, 2022

N T AB0 : Design Priors for AI-Augmented Generative Design of Network Tied-Arch-Bridges 5

are calibrated to the training set using the Logloss function for binary categorical properties and the
MultiClass loss function for categorical properties with more than two categories (i.e. classes). Model
evaluation was also conducted on the test data by calculating the accuracy and balanced accuracy.
For both algorithm heads, hyper-parameter tuning was performed using grid search.

4 Results

Figure 3: 2D cut scatter plot of the multi-dimensional clusters with a representative bridge for each

Figure 4: 2D scatter plot of the multi- Figure 5: 2D swarm plot of the four multi-dimensional
dimensional clusters (+: Mean value of clusters identified with the k-Prototype algorithm
cluster; 70% confidence interval ellipse) (25%- and 75%-quantile Boxplot limits).

For the clustering analysis a reasonable number of clusters for the k-Prototype algorithm is established
upon the resulting cost degradation with increasing number of clusters k, which shows a slope change
("elbow") at k = 4 for multiple γ-values. A suitable γ was found to be 4. The four multi-dimensional
clusters found in the bridge data set by the k-Prototype algorithm are visualised in Figs. 3, 4, and 5. A
good separation between the clusters is visible in all dimensions, where we show as an example the
distributions for span to rise, width-tie-back to span-to-rise ratio, and span to arch-bracing. These are
to be interpreted as 2D-cuts of the multi-dimensional cluster.

33. Forum Bauinformatik, München, 2022

6 Sophia V. Kuhn, Rafael Bischof, Georgios Klonaris, Walter Kaufmann, Michael A. Kraus

Table 1: Prediction performances of the regression - classification models of the N T AB0 prior model.

The achieved performance of the trained ML model w.r.t. individual parameter predictions are summa-
rized in Table 1. Initially, we assume the parameters span, tie back width and bridge function as fixed.
The remaining parameters are predicted in series by the ML model. For the classification heads, a
high predictive performance was achieved for all parameters on the training set. On the evaluation set
an accuracy of above 77% was achieved for all parameters, but a decreased balanced accuracy for
arch and tieback material, hanger arrangement and arch bracing is visible. The regression models are
performing well on the training set, but exhibit a substantially increased RMSE on the evaluation set.

5 Discussion

The k-Prototype appropriately solved the pattern recognition task given the difficulty of the mixed data
types present. Four clusters, distinctly well-separated in all dimensions, were found, which proves
the equal contribution of all bridge features (both continuous and categorical) during the clustering
process. Comparing the probability distributions of the individual features within the clusters reveals
interesting relations: The identified bridge clusters are clearly separated w.r.t the span, indicating
strong similarities in all bridge properties for the span ranges of the clusters (Fig. 3). Another dominant
distinction is the span-to-rise ratio, separating mostly pedestrian bridges from the road and rail bridges
(Fig. 4). Fig. 5 displays the cross girder as the most frequently used arch bracing type, mostly used
for medium-span NTABs. In addition, engineers in the past have found K-truss bracings beneficial for
short-span NTABs and diamond truss bracings effective for long-span NTABs. These detected distinct
patterns indicate a good chance for calibrating discriminative ML prediction models for bridge features.

Evaluation of the trained discriminative model showed that gradient boosted decision trees are suitable
to capture the dependencies of all bridge parameters in the training set. A reduced balanced accuracy
compared to the unbalanced accuracy is identified on the evaluation data set for the classification of
some of the bridge parameters. The reason here is the strong imbalance of classes present in the data
set, leading to a more performant ML classifier for the dominant classes. Improving the performance
by employing skew-insensitive metrics such as DKM and Hellinger distance as splitting criteria in the

33. Forum Bauinformatik, München, 2022

N T AB0 : Design Priors for AI-Augmented Generative Design of Network Tied-Arch-Bridges 7

construction of the decision tree [16] is omitted given the small size of the bridge data set. The limited
performance of the regression algorithm on the evaluation data is a result of overfitting, which is often
detected for small data sets. Another probable reason is that the available features are insufficient to
find a generalised relationship. While the performance evaluation proves that the ML model provides
suitable recommendations based on past bridge projects, we can also compare the recommendations
to research findings on NTABs. For an input parameter vector falling in the medium-span cluster (red in
Fig. 5) the model recommends a span-to-rise ratio of 6.4, which matches literature recommendations
of 5.8 to 6.67 [17]. For the same example, the model advocates the cross girder due to its frequent use,
while [18] identifies it to be the least cost-efficient arch bracing type. This reveals a limitation of the the
presented method. It does not evaluate weather the underlying data of existing bridges is compliant
with current design standards or good engineering practice w.r.t. efficiency, durability, etc.. Hence,
the method can adopt systematic mistakes from past bridge projects due to its data-driven nature.
Additionally, The cluster analysis and prior model calibration were performed on a small data set of
203 bridges and can therefore not claim to be entirely representative of all the already built NTABs
worldwide. Consequently, gathering a larger data set of NTABs is initiated for future investigations.
While the approach was applied to NTABs in the present work, the implemented framework is directly
applicable to further data sets of other frequently built structure types.

6 Conclusions

Today, the conceptual bridge design remains disconnected from performance evaluations of later
stages and therefore dictated by the prior knowledge of the involved experts, leading to a lengthy and
costly design process characterized by many iterations. We provide a two-stage method of clustering
a data set of structural bridge information and subsequent building discriminative regression resp.
classification models as informative design prior. The performed k-Prototype cluster analysis detected
design patterns for NTABs in the form of 4 distinct clusters. The identified structure enabled to draw
useful conclusions about sensible bridge parameter choices, which can be checked for plausibility by
bridge engineering experts yet also inform about hidden design patterns. The discriminative gradient
boosted decision tree algorithms serve as a powerful prior model for predicting sensible parameters
for a new bridge project situation based on existing bridge construction projects. The two-stage prior
modelling approach is found especially useful for detecting multi-dimensional dependencies between
the bridge features, which are not easily identifiable by conventional methods. Identified limitations are
the risk of adopting mistakes from the past and the limited availability of data in the AEC sector.

Acknowledgements

The authors would like to thankfully acknowledge the facilities of Design++ at ETH Zürich and the
funding through ETH Foundation grant No. 2020-HS-388 (provided by Kollbrunner/Rodio).

33. Forum Bauinformatik, München, 2022

 8 Sophia V. Kuhn, Rafael Bischof, Georgios Klonaris, Walter Kaufmann, Michael A. Kraus

References

[1] M. A. Kraus, “Machine learning techniques for the material parameter identification of laminated
glass in the intact and post-fracture state”, 2019.
[2] C. M. Bishop, Pattern Recognition and Machine Learning. Springer, 2016.
[3] B. Ercan and S. T. Elias-Ozkan, “Performance-based parametric design explorations: A method
for generating appropriate building components”, Design Studies, vol. 38, pp. 33–53, 2015.
[4] S. Burrows, B. Stein, J. Frochte, D. Wiesner, and K. Müller, “Simulation data mining for supporting
bridge design”, in Proceedings of the Ninth Australasian Data Mining Conference-Volume 121,
Citeseer, 2011, pp. 163–170.
[5] V. Ahmed, Z. Aziz, A. Tezel, and Z. Riaz, “Challenges and drivers for data mining in the aec
sector”, Engineering, Construction and Architectural Management, 2018.
[6] A. Müller, S. Kuhn, and M. Kraus, Scientific machine learning for structural engineering repository,
2022. [Online]. Available: https://sciml4structeng.github.io/Repository, (accessed: 07.07.2022).
[7] G. Batista and M.-C. Monard, “A study of k-nearest neighbour as an imputation method.”, vol. 30,
Jan. 2002, pp. 251–260.
[8] P. Tveit, “Considerations for design of network arches”, Journal of Structural Engineering,
vol. 113, no. 10, pp. 2189–2207, Oct. 1, 1987, Publisher: American Society of Civil Engineers.
[9] W. Kaufmann, Lecture Notes: Arch Bridges, 2021.
[10] N. de Vos. “Kmodes”. (2021), [Online]. Available: https://datasolut.com/wiki/clusteranalyse/.
(accessed: 10.04.2021).
[11] E. W. Forgy, “Cluster analysis of multivariate data : Efficiency versus interpretability of classifica-
tions”, Biometrics, vol. 21, pp. 768–769, 1965.
[12] Z. Huang, “Extensions to the k-means algorithm for clustering large data sets with categorical
values”, Data Mining and Knowledge Discovery, pp. 283–304, 1998.
[13] Andrewngai, “Understanding dbscan algorithm and implementation from scratch”, Towards Data
Science, 2020.
[14] Y. LLC, Catboost, 2019. [Online]. Available: https://github.com/catboost/catboost.
[15] A. Jha. “CatBoost – A new game of Machine Learning”. (2020), [Online]. Available: https :
//affine.ai/catboost-a-new-game-of-machine-learning/. (accessed: 27.06.2021).
[16] W. Daelemans and B. Goethals and K.Morik (Eds.), Machine Learning and Knowledge Discovery
in Databases - ECML PKDD 2008, Part I. Springer Berlin Heidelberg, 2008, pp. 241–256.
[17] P. Tveit, “Systematic Thesis on Network Arches”, Adger University, 2014.
[18] F. Schanack, “Network arch bridges”, Ph.D. dissertation, University of Cantabria, 2008.

33. Forum Bauinformatik, München, 2022

View publication stats