Artificial Intelligence: An Aid in Structural Engineering

Due to the unstructured, inexact, and iterative structural engineering process that requires the
skills gained from experience like defining the system, generating alternatives, evaluating
alternatives, and decision-making, researchers have investigated the potential of artificial
intelligence as an alternative approach and complement to the solution of structural engineering
problems. The main idea of this paper is to demonstrate why and how artificial intelligence has
been used to solve different problems encountered in structural engineering. Note, however,
that this paper does not list all the studies on the applications of artificial intelligence in structural
engineering. The reader is referred to Furuta, Tu and Yao (1985), Adeli (1986), Adeli (2001),
Amezquita-Sanchez et al. (2016), Salehi and Burgueño (2018), and Sun, Burton and Huang
(2020) for a more complete list of studies. Rather, this will focus on showing the different
challenges in structural engineering in the course of time and how artificial intelligence methods
help in solving these challenges. To do this, the structural engineering field will be described in
parallel with the eras of artificial intelligence: Pre-AI era, Symbolic AI era, and the Data-Driven
Machine Learning era.

Developments in computer technology opened new ways to do structural engineering.

Engineers have integrated the use of computers in the engineering process to produce more
accurate outputs in a shorter amount of time. Since the 1950s, Computer-Aided Design (CAD)
has been used as a term to integrate using computers into the design process. In 1957, Patrick
Hanratty developed a numerical-control programming system to automate machine tools which
started the development of CAD (Myers, 1998). In 1963, Ivan Sutherland created Sketchpad
which aimed to replace the pencil and paper method used to create structural design drawings.
Sketchpad pioneered the development of human-computer interaction by simplifying the
process of drawing geometry through a graphical user interface. This solves the problem of
repeating structural drawings created with pencil and paper by a draftsman due to the iterative
nature of the design.

In addition to advances in computer graphics, improvements in numerical analysis of structures

also flourished. Due to the improved computing power and the availability of microcomputers
and personalized computers, structural analysis methods like the finite element method were
easier to integrate into CAD systems and more accessible for engineers (Knowles, 1984). This
reduced the risk of human error and analysis time compared to traditional manual calculations,
resulting in a better and more cost-effective design.
However, users of computer analysis programs were faced with the question of the accuracy of
the models they generated. Experts are needed to define the system, offer design alternatives
as input into computer software, interpret the numerical models, assess whether numerical
models reflect the actual scenario, and decide the best alternative for the design. (Masters,
1993) argued that computers cannot make decisions because they only execute logical
instructions provided to them. He claimed it was impossible to define the design process as
logical instructions that a computer could execute. Human experts are better at making
decisions because they have developed their engineering intuition learned from previous
experiences. Engineering intuition and the thought process of an expert are not logically
defined, but an art.

In relatively simpler problems, structural optimization programs were developed to simulate the
decision-making of an expert (Radley, 1969; Ellis and West, 1970; Pitts, 1970). These
optimization methods aim to automatically find the configuration of a structural member defined
by design variables including cross-sectional area of the member, strength of concrete, and
number of reinforcements while satisfying design constraints like space limitations, cost, and
availability of materials. However, a human expert is still needed to provide the initial feasible
design that satisfies the constraints. Moreover, when the initial design is poor, these methods
tend to converge in a local optimum instead of a global optimum. Also, these methods do not
consider that the design process is dynamic, i.e., design variables and constraints may change.

Concurrently, researchers explored how to build intelligent machines that simulate human
intelligence. In 1956, computer scientists led by Marvin Minsky and John McCarthy coined the
term Artificial Intelligence in a summer workshop in Dartmouth College (Buchanan, 2005). This
was followed by significant projects on building systems which mimic an expert in solving a
problem called Expert Systems (Haenlein and Kaplan, 2019). Expert Systems solves problems
by reasoning through if-then rules and symbolic logic rather than sequential code. Also, a user
of expert systems can update the rules as needed. Researchers realized the advantage of using
expert systems to model engineering intuition gained through previous experiences represented
by explicit if-then rules. These systems can be represented as a set of inference questions
which outputs an evaluation of structural safety and damage state (Furuta, Tu and Yao, 1985).

However, expert systems have been criticized because it performs poorly on problems that do
not exhibit formalized rules. This applies to engineering design, which makes expert systems
impractical to use in actual design situations (Rooney, 1983). These systems tend to be
restricted to simple problems. Moreover, there are simply too many rules to consider and some
of them conflict for different structural systems.

This led to another approach to artificial intelligence. Instead of representing intelligence as rule-
based, the learning problem was patterned to how the brain works. In 1958, Rosenblatt
introduced the concept of a perceptron which will become the basis of artificial neural networks.
A perceptron unit accepts an input, processes it, and outputs a value of zero or one. It mimics
the behavior of a neuron when it accepts signals from stimuli and decides whether to fire up or
not. Perceptrons can be built in layers to form a network which can be optimized to map known
inputs and outputs it learns from examples. One of the advantages of perceptron learning is that
it is data-driven. Users do not have to update the code with rules. Instead, the parameters in the
computation of the output of the perceptron is adjusted when the predicted output is not the
same as the expected output. It can be argued that these adjustments, without the need for
changing the code architecture, signify true learning by examples.

In the structural engineering field, the ability of the perceptron to gain experience through self-
adjustment was utilized. Adeli and Yeh (1989) applied perceptron learning in a steel beam
design using steel beams that satisfy the design requirements of the American Institute of Steel
Construction (AISC) -Load and Resistance Factor Design specification as training data. In their
example, it was noted that the degree of precision of the perceptron network increases as the
training time increases. However, perceptron learning requires a large memory space, faster
and more efficient computing machinery, and many training examples to learn. Perceptron
learning evolved into Artificial Neural Networks as methods for optimizing these networks
improved. Following the research of Adeli and Yeh, many studies aimed to explore using
artificial neural networks in different problems in civil engineering. The pattern recognition
capability of ANN was demonstrated in locating the load applied in a structural member and in
determining the required depth of a concrete beam given the design constraints like the bending
moment, yield strength, compressive strength, and reinforcing ratio (VanLuchene and Sun,
1990). Messner, Sanvido and Kumara (1994) used ANN to get a suggestion for a preliminary
design section which lessens design iterations. Hajela and Berke (1991) noted the following
advantages of neural networks over the traditional numerical computing paradigm. Once the
neural network is trained, it can rapidly map the inputs and outputs which speeds up the design
process. This is noteworthy when nonlinearities are present in the analysis. Neural networks
can also abstract and extract core concepts with virtually no constraint in representation. This is
opposed to using expert systems which heavily requires for the expertise of the user to improve
the learning. Several studies used ANN to learn from experimental and heuristic data (Adeli,
2001).

Elkordy, Chang, and Lee (1994) pointed out that traditional methods for structural system
identification and health monitoring based on visual inspections and on-site tests were
unreliable in producing accurate representations of structures due to considerable noise and
incomplete datasets. Many studies proposed to use artificial neural networks trained on
combined experimental results from the shaking table test and numerical analyses to estimate
the stiffness parameters of two-dimensional structures (Yun and Bahng, 2000) and three-
dimensional structures (Huang and Loh, 2003). Stephens and VanLuchene (1994) concluded
that neural networks can simplify the assessment of structures by reducing several indicators
into a reliable single indicator. Several studies at the start of the 20th century used ANN in
detecting and locating damage in structures (Amezquita-Sanchez et al., 2016).

The research on active control of structures also grew at the last decade of the 20th century.
Ghaboussi and Joghataie (1995) demonstrated the potential of neural networks to control the
structure in a simulation. They called the system as a neurocontroller which outputs an actuator
force given the structural response represented by signals measured by sensors. Ankireddi and
Yang (1999) used artificial neural networks to detect failure by detecting inconsistency in the
signals received by a sensor measuring the dynamic properties of the structure.

Although numerical methods like the Finite Element Method were improved, the accuracy of the
model still heavily relies on the user. For example, the choice of the number of nodes, number
of elements, shape function, and time step affect the accuracy of the numerical model. Studies
proposed to use artificial neural networks to aid in choosing the correct parameters for
numerical analysis (Adeli, 2001).

Researchers also explored other machine learning methods not based on neurons like
regression methods, decision trees, nearest neighbors, support vector machines, discriminant
analyses, and evolutionary algorithms in damage detection, structural health monitoring, and
design (Salehi and Burgueño, 2018; Sun, Burton and Huang, 2020). These methods, like neural
networks, can learn by using data examples without prior knowledge on the theory about the
structural system.

However, the main problems in using machine learning methods, including artificial neural
networks, is the lack of interpretability and lack of data. In general, artificial intelligence
researchers found that using learning models developed through bottom-up optimization like
neural networks in different applications is impractical. First, a huge amount of data is needed to
accurate fit the results of a learning architecture to the expected output. This is a problem
because obtaining example data which includes the input and the corresponding output
especially for scientific problems are difficult to obtain. Also, machine learning methods exhibits
a black-box behavior which makes the users hesitant to fully embrace these methods because
they cannot interpret what these methods are doing. Regarding the structural engineering field,
Hung et al., (2003) noted that ANN cannot perform well on more complex and nonlinear
structural systems because of these disadvantages.

To solve the lack of interpretability, researchers have explored the idea of combining expert
systems and neural networks. The goal of this combination is to produce more robust learning
architecture by using the logical rules of expert systems as a guide to the learning of neural
networks. This would later be called Guided Machine Learning. Adeli and Jiang (2006)
combined fuzzy expert system and neural network to perform structural identification more
efficient when the behavior is nonlinear and when a significant amount of noise is present in the
measurements. This was followed by similar studies that combine expert systems and neural
networks applied in different structural engineering applications like identifying damping
properties, determining natural frequencies, damage identification, and design. Zhang and Sun
(2020) proposed a modified loss function to reflect the discrepancy between the prediction and
expected output that consistent with physics.

Overfitting is a major problem in machine learning. In structural engineering problems, this is

aggravated by insufficient high-quality data since it is difficult and costly to obtain. Sun, Burton
and Huang (2020) called to have a unified effort for researchers to develop a benchmark data
for structural engineering and to explore the combination of data-driven machine learning
models and theory to tackle this problem.

Another way to tackle the problem on insufficient data is to use semi-supervised learning. In a
supervised learning, the machine is trained using both labeled and unlabeled data. This is
advantageous since in many structural engineering problems, a vast amount of input data can
be generated. Since evaluating the effects of these input data is too costly through experiments
and simulations, only a small percentage of the input data is evaluated to produce the
corresponding label output. Semi-supervised learning was used in the study of Zhang and Sun
(2020) using labeled and unlabeled simulation data.
With the advances in computer technology, the field of structural engineering has also
developed. The development of computer hardware, computer graphics, and numerical
computing methods made engineering analysis and design faster and more efficient. However,
the accuracy of the numerical models heavily relied on the expertise of the user gained through
previous experience. Researchers have attempted to simulate how engineering experts think
using expert systems. They represented decision-making using symbolic logic, which worked
well for relatively simpler problems. However, for more complex systems, using expert systems
is impractical because too many rules are needed to be integrated into the system and some of
these rules conflict. This led to the exploration of data-driven learning models. Using these
models is advantageous because they do not need underlying theories and rules to learn.
Instead, they learn by example. Several learning models were developed in different structural
engineering problems including structural identification, structural health monitoring, active
control of structures, finite element analysis, and design automation. However, using purely
data-driven models in structural engineering problems has two disadvantages: insufficient data
and lack of interpretability. Current studies on the application of artificial intelligence in structural
engineering are geared to solve these problems. To solve the insufficiency of data, researchers
are developing accurate baseline datasets to be used for learning. Semi-supervised learning is
explored where many unlabeled building data in combination with few labeled data are used for
learning. The combination of rule-based and data-based learning is also currently investigated
to develop more interpretable learning models for structural engineering problems. In this article,
it was shown how the structural engineering field utilized the technological advances in the past
and how advances in artificial intelligence presented significant potential in improving the
structural engineering process in the future.

References:

Adeli, H. (1986) “Artificial intelligence in structural engineering,” Engineering Analysis, 3(3), pp. 154–160. doi:
10.1016/0264-682X(86)90053-5.

Adeli, H. (2001) “Neural networks in civil engineering: 1989-2000,” Computer-Aided Civil and Infrastructure
Engineering, 16(2), pp. 126–142. doi: 10.1111/0885-9507.00219.

Adeli, H. and Jiang, X. (2006) “Dynamic fuzzy wavelet neural network model for structural system identification,”
Journal of Structural Engineering, 132(1), pp. 102–111. doi: 10.1061/(ASCE)0733-9445(2006)132:1(102).

Adeli, H. and Yeh, C. (1989) “Perceptron Learning in Engineering Design,” Computer‐Aided Civil and Infrastructure
Engineering, 4(4), pp. 247–256. doi: 10.1111/j.1467-8667.1989.tb00026.x.
Amezquita-Sanchez, J. P. et al. (2016) “Neurocomputing in civil infrastructure,” Scientia Iranica, 23(6), pp. 2417–
2428. doi: 10.24200/sci.2016.2301.

Ankireddi, S. and Yang, H. (1999) “Neural networks for sensor fault correction in structural control,” Journal of
Structural Engineering, 1(2), pp. 1–11.

Buchanan, B. G. (2005) “A (very) brief history of artificial intelligence,” AI Magazine, 26(4), pp. 53–60.

Elkordy, M. F., Chang, K. C. and Lee, G. C. (1994) “A Structural Damage Neural Network Monitoring System,”
Computer‐Aided Civil and Infrastructure Engineering, 9(2), pp. 83–96. doi: 10.1111/j.1467-8667.1994.tb00364.x.

Ellis, J. A. and West, P. E. (1970) “A structural optimisation technique,” Computer-Aided Design, pp. 24–28.

Furuta, H., Tu, K. S. and Yao, J. T. P. (1985) “Structural engineering applications of expert systems,” Computer-
Aided Design, 17(9), pp. 410–419. doi: 10.1016/0010-4485(85)90288-X.

Ghaboussi, J. and Joghataie, A. (1995) “Active Control of Structures Using Neural Networks,” Journal of Engineering
Mechanics, 121(4), pp. 555–567.

Haenlein, M. and Kaplan, A. (2019) “A brief history of artificial intelligence: On the past, present, and future of artificial
intelligence,” California Management Review, 61(4), pp. 5–14. doi: 10.1177/0008125619864925.

Hajela, P. and Berke, L. (1991) “Neurobiological computational models in structural analysis and design,” Computers
and Structures, 41(4), pp. 657–667. doi: 10.1016/0045-7949(91)90178-O.

Huang, C.-C. and Loh, C.-H. (2003) “Nonlinear Identification of Dynamic Systems Using Neural Networks,”
Proceedings of the International Symposium on Test and Measurement, 2, pp. 997–1000. doi: 10.1016/b978-
012526430-3/50007-6.

Hung, S. L. et al. (2003) “Nonparametric identification of a building structure from experimental data using wavelet
neural network,” Computer-Aided Civil and Infrastructure Engineering, 18(5), pp. 356–368. doi: 10.1111/1467-
8667.t01-1-00313.

Knowles, N. C. (1984) “Finite element analysis,” Computer-Aided Design, 16(3), pp. 134–140. doi: 10.1016/0010-
4485(84)90036-8.

Masters, E. H. (1993) “The engineering design functions which the computer can most effectively aid,” Computer-
Aided Design, 25(2), pp. 130–135. doi: 10.1016/0010-4485(93)90098-9.

Messner, J. I., Sanvido, V. E. and Kumara, S. R. T. (1994) “StructNet: A Neural Network for Structural System
Selection,” Computer‐Aided Civil and Infrastructure Engineering, 9(2), pp. 109–118. doi: 10.1111/j.1467-
8667.1994.tb00366.x.

Myers, B. (1998) “A brief history of human-computer interaction technology,” Interactions, pp. 44–54. doi:
10.1002/(SICI)1099-1492(199610)9:7<279::AID-NBM435>3.0.CO;2-V.

Pitts, G. (1970) “Computer-aided design: use and misuse,” Computer-Aided Design, 2(4), pp. 41–45. doi:
10.1016/0010-4485(70)90017-5.
Radley, D. (1969) “Computer-aided design in the USA,” Computer-Aided Design, pp. 5–8.

Rooney, M. F. (1983) “Artificial intelligence in simple beam design,” Journal of Structural Engineering (United States),
109(9), p. 2225. doi: 10.1061/(ASCE)0733-9445(1983)109:9(2225.2).

Rosenblatt, F. (1958) “The perceptron: A probabilistic model for information storage and organization in the brain,”
Psychological Review, 65(6), pp. 386–408. doi: 10.1037/h0042519.

Salehi, H. and Burgueño, R. (2018) “Emerging artificial intelligence methods in structural engineering,” Engineering
Structures, 171(November 2017), pp. 170–189. doi: 10.1016/j.engstruct.2018.05.084.

Stephens, J. E. and VanLuchene, R. D. (1994) “Integrated Assessment of Seismic Damage in Structures,”

Computer‐Aided Civil and Infrastructure Engineering, 9(2), pp. 119–128. doi: 10.1111/j.1467-8667.1994.tb00367.x.

Sun, H., Burton, H. v. and Huang, H. (2020) “Machine Learning Applications for Building Structural Design and
Performance Assessment: State-of-the-Art Review,” Journal of Building Engineering, 33(July 2020), p. 101816. doi:
10.1016/j.jobe.2020.101816.

Sutherland, I. (1963) “Sketchpad: a man-machine graphical communication system,” in Spring Joint Computer
Conference. doi: 10.1111/j.1447-073x.2005.00096.x.

VanLuchene, R. D. and Sun, R. (1990) “Neural Networks in Structural Engineering,” Computer‐Aided Civil and
Infrastructure Engineering, 5(3), pp. 207–215. doi: 10.1111/j.1467-8667.1990.tb00377.x.

Yun, C. B. and Bahng, E. Y. (2000) “Substructural identification using neural networks,” Computers and Structures.
doi: 10.1016/S0045-7949(99)00199-6.

Zhang, Z. and Sun, C. (2020) “Structural damage identification via physics-guided machine learning: a methodology
integrating pattern recognition with finite element model updating,” Structural Health Monitoring. doi:
10.1177/1475921720927488.