Smart Structures

Implementing Metaheuristics and Machine

Learning in Structural Engineering
September 13th, 2025

Speaker:
Santiago Londoño López
MSc S.E., C.E.

[Link]@[Link]

Politecnico di Torino
Corso Duca degli Abruzzi, 24. Torino, Piemonte, Italia.
 Content
• AI further than Chatbot Applications – Generative Prompts

• Application 1: Structural Isolation Devices – ArtIStE

• Application 2: Directional Earthquakes – ArtIStE & RealSmart

• Application 3: Exoskeletons - ArtIStE

• Application 4: Bridge Damage Identification - RealSmart

Santiago Londoño López

Politecnico di Torino
DISEG – ArtIStE laboratory
AI – Further than Chatbot Applications??

2015-2025
• Buildings: roughly 500-1,000 collapse events globally that
are significant over the last10 years.
• Bridges: around 200-500 bridge collapse events worldwide
over the same period. Santiago Londoño López
Politecnico di Torino
DISEG – ArtIStE laboratory
AI – Further than Chatbot Applications??

Evolutionary
Algorithms

Santiago Londoño López

Politecnico di Torino
DISEG – ArtIStE laboratory
 Operative Definition
• Optimization as searching in an intelligent manner (Luenberger)
• Optimization techniques stem from variational calculus where maximum or minimum
of a function is sought
• Several applied sciences have been using the same mathematical concepts:
• optimal design in engineering with starting applications in aeronautics
• operational research in economics and management social science

Santiago Londoño López

Politecnico di Torino
DISEG – ArtIStE laboratory
General WorkFlow
x = x0

evaluate f (x)

no
convergence? new x

yes xopt = x ⇤
Santiago Londoño López
end Politecnico di Torino
DISEG – ArtIStE laboratory
 Operative Definition
Function to be optimized can be single f (X) (Objective
Function OF , cost function etc.) or can be made by an array f
of functions:
• single function ! scalar optimization
x
• functions array ! vector optimization or multi-
objective
OF can be evaluated in different ways:
• analytical
• numerical (FEM code, circuit simulation etc.) f
approximated by surrogate models
(interpolation, neural network etc.)
x
Santiago Londoño López
Politecnico di Torino
DISEG – ArtIStE laboratory
Multimodal functions allow two kinds of stationary
points:
• local optimum: point that extremates the
function in a local set local global
• global optimum: point that extremates the
function in the whole feasibility region f
Scalar optimization problems can be subdivided in classes
depending on:
• objective function characteristics
x
• d.o.f number x x
• optimization algorithm L U

Santiago Londoño López

Politecnico di Torino
DISEG – ArtIStE laboratory
 Optimization Algorithms
No rule can be defined to pick up the best algorithm so that a choice must be
made looking for a trade-off between:
• convergence speed
• reliability of the solution
Algorithms can be subdivided according to their strategy toward optimum:
• classic or deterministic algorithms
• stochastic algorithms - Metaheuristic
Another classification can be made according to the number of dof:
• small case large scale

Santiago Londoño López

Politecnico di Torino
DISEG – ArtIStE laboratory
 Deterministic Algorithms
Deterministic algorithms look for the minimum of a function using variational
concepts
• null gradient
• second derivative or Hessian matrix positive definite (minimum)
Quality of result depends on starting point of search
Speed of the method is function of the way in which gradient/Hessian is computed
(analytical/numerical)
Modified formulations of the problem (Lagrangian) allow to treat constraints in an exact way

Santiago Londoño López

Politecnico di Torino
DISEG – ArtIStE laboratory
 Deterministic Algorithms
Classification can be made according to the derivation order of OF:
• Zero-th order methods use only sampled values of the OF
• First order methods use sampled values of the OF and of its gradient
• Second order methods use sampled values of the OF, gradient and Hessian
If OF is known only in numeric way and if n is the number of d.o.f, evaluation of
gradient require computation of the function in n + 1 points

Santiago Londoño López

Politecnico di Torino
DISEG – ArtIStE laboratory
 Stochastic Algorithms
Not withstanding large differences between algorithm characteristics, some common
topics can be highlighted :
• they are inspired by some natural problem solving
• they are independent on the problem
• they are all belonging to zero-th order algorithms

Santiago Londoño López

Politecnico di Torino
DISEG – ArtIStE laboratory
 Stochastic Algorithms
All algorithms try to find the global optimum of the OF and they use
some common rules:
• use of a random factor that makes each trajectory toward
the optimum different run by run
• generally, they accept a deterioration of the OF
• accumulation of small improvements
As a consequence:
• algorithms converge slowly toward global optimum
• evaluation of their performances must be made in a
statistical way

Santiago Londoño López

Politecnico di Torino
DISEG – ArtIStE laboratory
 Stochastic Algorithms
Usually, success of optimization is a trade-off between:
• Exploration: try to find which part of the feasibility region is most promising
• Exploitation: try to reach the optimum as fast as possible
two kinds of search are extremal in this sense:
• Random search: maximum exploration since every new configuration is independent
from results got in the past
• Hill-climbing: local exploration in the nearby of a good point accepting only up-hill
moves

Santiago Londoño López

Politecnico di Torino
DISEG – ArtIStE laboratory
 Stochastic Algorithms
If a method is able to reach often a good trade-off between exploration and exploitation it
Is called robust (Michalewicz)
robustness is a quality that:
• leads to quick convergence in simple problems
• is able to try different regions and is not trapped in local minima in hard problems
a domain independent algorithm is supposed to tackle both simple and hard problems
without any parameters tuning

Santiago Londoño López

Politecnico di Torino
DISEG – ArtIStE laboratory
 Stochastic Algorithms
On the contrary, classic algorithms have a serial structure
this issue makes the two classes of methods complementary:
• classic algorithms: serial and with quick convergence to the closest optimum
(exploitation)
• stochastic algorithms: parallel and with good exploration properties
Hybrid techniques can be defined with a stochastic algorithm at the control level
and with a classic one taking care of the basic level combining so their
characteristics.

Santiago Londoño López

Politecnico di Torino
DISEG – ArtIStE laboratory
GENETIC ALGORITHM PARTICLE SWARM
OPTIMIZATION
POPULATION
INITIALIZATION POPULATION
INITIALIZATION

PARENT
SELECTION
UPDATE
POSITIONS
CROSSOVER
no

MUTATION end FITNESS

no criteria? EVALUATION

FITNESS end yes

EVALUATION criteria?
OPTIMAL
yes SOLUTION
OPTIMAL
SOLUTION

Stochastic Algorithms
Santiago Londoño López
Politecnico di Torino
DISEG – ArtIStE laboratory
 Optimization Definition
Function to be Optimized Parameters
𝐹 𝑥, 𝑦 = sin 0.02 ∗ 𝑥 ∗ cos 0.05 ∗ 𝑦 +

0.2 ∗ sin 0.1 ∗ 𝑥 ∗ cos 0.2 ∗ 𝑦 +

0.3 ∗ 𝑒
− 𝑥2 +𝑦2
100 ∗ 𝑠𝑒𝑛 0.2 ∗ 𝑥 ∗ cos(0.2 ∗ 𝑦)
Number of
Runs 𝟏𝟎
Population
Size 𝟏𝟓
Number of
Iterations 𝟑𝟎𝟎
Santiago Londoño López
Politecnico di Torino
DISEG – ArtIStE laboratory
 Genetic Algorithm (G.A)
PARENT Roulette Uniform
SELECTION CROSSOVER Scattered MUTATION
Wheel prob.: 0.01

FOUND THE
OPTIMAL 𝟏𝟎ൗ FOCUSED ON EXPLORATION
Santiago Londoño López
Politecnico di Torino
SOLUTION 𝟏𝟎 DISEG – ArtIStE laboratory
 Particle Swarm Optimization (P.S.O)
OWN SOCIAL social adj. weight self adj. weight
VELOCITY inertia GLOBAL BEST
SELF BEST

FOUND THE
OPTIMAL 𝟕ൗ FOCUSED ON EXPLOITATION Santiago Londoño López

SOLUTION 𝟏𝟎 Politecnico di Torino

DISEG – ArtIStE laboratory
AI – Further than Chatbot Applications??
Feature Deterministic Algorithms Stochastic Algorithms

Predictable, same output for same

Behavior Randomized, may yield different results per run
input
Convex, smooth, well-defined
Best for Noisy, multimodal, high-dimensional problems
problems

Gradient Descent, Newton-Raphson

Examples Genetic Algorithms, PSO, Simulated Annealing
’s Method

Speed Often faster in structured problems Slower but more exploratory

Robustness Sensitive to initial conditions Better at escaping local minima

Mathematically elegant, but limited in

Complexity Heuristically rich, often hybridized
scope

Santiago Londoño López

Politecnico di Torino
DISEG – ArtIStE laboratory
 Case of Study 1
High Damping Rubber
Bearings: Optimal models
evaluation considering
production variability

(Londono-Lopez, et al, 2023) - ArtIStE

Contextualization:

Santiago Londoño López

Politecnico di Torino
DISEG – ArtIStE laboratory
Device Description :
Description Unit Value
Max. Vertical Load (ULS) FUSL [kN] 2400
Max. Vertical Load (CLS) FCLS [kN] 1500
Max. deformation (CLS) dED [mm] (±)250
Equivalent Damping ξ% 16%
Num. Rubber Layers [-] 26
Thickness of the Rubber Layer [mm] 8

Santiago Londoño López

Politecnico di Torino
DISEG – ArtIStE laboratory
• 3 Production Lots [2021-2023]
Description of laboratory test:
• 160 HDRB ISI-S 600/208 P40
series
• Signal cleaning FFT – Moving
average.

Santiago Londoño López

Politecnico di Torino
DISEG – ArtIStE laboratory
Numerical Laws:

Santiago Londoño López

Politecnico di Torino
DISEG – ArtIStE laboratory
 Identification Framework :
The parameter identification requires solving an unconstrained optimization problem where a suitable
objective function (OF) (W) must be minimized. Using the preset MATLAB Genetic Algorithm.

𝑥1𝑙
…
𝑙
𝑥𝑖𝑙 = 𝑥𝑗
…
𝑥𝑛𝑙

𝑥1𝑢
Number of Number of …
Population Size Evaluation Stagnation 𝑢
Generations Runs 𝑥𝑖𝑢 = 𝑥𝑗
100 50 5000 20 0.05 …
𝑥𝑛𝑢

Santiago Londoño López

Politecnico di Torino
DISEG – ArtIStE laboratory
 Jankowski Model:
A non-linear strain rate-dependent model designed for HDRB devices subjected to a constant vertical load
and invariant temperature is assumed. It simulates the horizontal behavior of the device considering a
nonlinear elastic spring dashpot element. The shear stiffness, K, and the damping coefficients, C, evaluated
at a given time t.

R. Jankowski, Non-linear model of high damping rubber bearing (2004).

Santiago Londoño López

Politecnico di Torino
DISEG – ArtIStE laboratory
Calibration through GA Optimization
Results:
Comparison between experimental and numerical force-displacement relationships for the algebraic model

OF mean (µ) considering 20 independent runs for all

devices: Santiago Londoño López
𝑂𝐹 = 8,10% Politecnico di Torino
DISEG – ArtIStE laboratory
 Vaina Model:
Vaina’s Complex Asymmetric mechanical hysteresis model introduces the evaluation of the generalized force
by solving a scalar equation and employing only one history-dependent variable. The model aims to resolve
two specific instances of the system: first, targeting the asymmetric bilinear trend, and second, the
exponential trend. To compute the model generalized force, four curves are calculated aiming to characterize
the shape of a standard asymmetric hysteresis loop.

N. Vaiana, S. Sessa, L. Rosati, A generalized class of uniaxial rate-independent

models for simulating asymmetric mechanical hysteresis phenomena, Mechanical
Systems and Signal Processing 146 (2021) 106984.

Santiago Londoño López

Politecnico di Torino
DISEG – ArtIStE laboratory
 Vaina Model:

N. Vaiana, S. Sessa, L. Rosati, A generalized class of uniaxial rate-independent

models for simulating asymmetric mechanical hysteresis phenomena, Mechanical
Systems and Signal Processing 146 (2021) 106984.

N. Vaiana, S. Sessa, L. Rosati, A generalized class of uniaxial rate-independent

models for simulating asymmetric mechanical hysteresis phenomena, Mechanical
Systems and Signal Processing 146 (2021) 106984.
Santiago Londoño López
Politecnico di Torino
DISEG – ArtIStE laboratory
Calibration through GA Optimization
Results:
Comparison between experimental and numerical force-displacement relationships for the algebraic model

OF mean (µ) considering 20 independent runs for all

devices: Santiago Londoño López
𝑂𝐹 = 8.76% Politecnico di Torino
DISEG – ArtIStE laboratory
 Abe et al. Model:
This model incorporates three equations that together contribute to the evaluation of the horizontal force. The
first component involves an elastic nonlinear spring F1, the second component introduces the hysteretic
force in the form of an elastic-plastic spring F2 while the third comes from a second elastic nonlinear force
that accounts for hardening F3:

M. Abe, J. Yoshida, Y. Fujino, Multiaxial behaviors of laminated rubber bearings

and their modeling. i: Experimental study, Journal of Structural Engineering 130
(8) (2004) 1119–1132.
Santiago Londoño López
Politecnico di Torino
DISEG – ArtIStE laboratory
Calibration through GA Optimization
Results:
Comparison between experimental and numerical force-displacement relationships for the algebraic model

OF mean (µ) considering 20 independent runs for all

devices: Santiago Londoño López
𝑂𝐹 = 8.48% Politecnico di Torino
DISEG – ArtIStE laboratory
 Hwang and Wu Model:
In this model, the shear force acting on the bearing is considered as the sum of the restoring force and the
damping force. In this context, F is the shear force transmitted by the bearing, which consists of the restoring
force F1 and the damping force F2. K(x(t);x′(t)) represents the stiffness of the bearings, while C(x(t);x′(t)),
is the damping coefficient. Both stiffness and damping coefficient depend on the relative displacement (x(t))
and Velocity (x´(t)) input.

J. Hwang, J. Wu, T.-C. Pan, G. Yang, A mathematical hysteretic model for

elastomeric
isolation bearings, Earthquake engineering & structural dynamics 31 (4) (2002)
771–789. Santiago Londoño López
Politecnico di Torino
DISEG – ArtIStE laboratory
Calibration through GA Optimization
Results:
Comparison between experimental and numerical force-displacement relationships for the algebraic model

OF mean (µ) considering 20 independent runs for all

devices: Santiago Londoño López
𝑂𝐹 = 11.68% Politecnico di Torino
DISEG – ArtIStE laboratory
Accuracy Algebraic Models -
Statistical Results:

Santiago Londoño López

Politecnico di Torino
DISEG – ArtIStE laboratory
Accuracy Differential Models -
Statistical Results:

Santiago Londoño López

Politecnico di Torino
DISEG – ArtIStE laboratory
 Bi Linear Parameter’s Statistical
Results:
• 3 Parameters model

Santiago Londoño López

Politecnico di Torino
DISEG – ArtIStE laboratory
Jankowski Parameter’s Statistical Results:

Regulates stiffness near maximum displacements Modify the degree of damping to shape the loops as desired

• 11 Parameters model Santiago Londoño López

Politecnico di Torino
DISEG – ArtIStE laboratory
Hwang and Wu Parameter’s Statistical
Results:
• 10 Parameters model

Parameter a10 represents the

energy dissipated by the
elastomer during cyclic
loading helps to manage the
reduction in loop area.

Santiago Londoño López

Politecnico di Torino
DISEG – ArtIStE laboratory
Closing
Remarks:
 Case of Study 2
Directional Variability ff
Ground Motion
Components: Toward an
Integrated Approach for
Seismic Hazard Assessment
(Londono-Lopez, et al, 2024) – ArtIStE & RealSmart
 Contextualization and Problem Statement

AFP. (2023, February 20). Collapsed buildings in Antakya, southern Turkey after a 6.4-magnitude earthquake TOI. (2025, March 28). Myanmar-Thailand earthquake: Videos capture collapsing buildings, cracked roads, and mass panic
[Photograph]. Hindustan Times. [Link] [Video. The Time of India. 7.7 Magnitude Earthquake. [Link]
[Link] earthquake-videos-capture-collapsing-buildings-cracked-roads-and-mass-panic/ar-
AA1BPCrJ?apiversion=v2&noservercache=1&domshim=1&renderwebcomponents=1&wcseo=1&batchservertelemetry=1
&noservertelemetry=1

Santiago Londoño López

Politecnico di Torino
DISEG – ArtIStE laboratory
 Fault Typology Interested Region:
Triggering mechanisms are,
at present, remain
effectively unquantifiable
and essentially
unpredictable:

• Rupture Directivity
• Fault Segmentation
• Asymmetric Sediment
Layers
Figure 1. (A) Simplified tectonic • Topographic
map of Italy and surrounding
• …
regions.

Figure extracted from: Billi, A., Gambini, R., Nicolai,

C., & Storti, F. (2007). Neogene-Quaternary Santiago Londoño López
intraforeland transpression along a Mesozoic platform-
basin margin: The Gargano fault system, Adria, Politecnico di Torino
Italy. Geosphere, 3(1), 1-15.
DISEG – ArtIStE laboratory
 Design Paradigm RSA -
NLTHA:

Chopra, A.K., 1995. Dynamics of structures theory and applications

of earthquake engineering
Khan, M.A., 2013. Earthquake-Resistant Structures: Design, Build,
and Retrofit. Butterworth-Heinemann.

Santiago Londoño López

Politecnico di Torino
DISEG – ArtIStE laboratory
 Geographic Location of Dataset Epicenters

2,809 Earthquake recordings

provided by the National
Institute of Geophysics and
Volcanology (INGV), Italy.
+15 Year window of
exploration.

Santiago Londoño López

Politecnico di Torino
DISEG – ArtIStE laboratory
 Data
Processing.

Polar Reference System (Ro)

Rotated and Normalized

Santiago Londoño López

Politecnico di Torino
DISEG – ArtIStE laboratory
 Data
Processing
.

Kernel Density Estimation

Circular Envelope
Convex Hull
Ellipsoidal b-Value

Santiago Londoño López

Politecnico di Torino
DISEG – ArtIStE laboratory
 K-Means Clustering on Seismogenetic characteristics
Cost Function (J)

✓ 𝑲: Number of clusters
✓ 𝒙𝒊 : A data point in the dataset
✓ 𝑪𝒌 : The set of data points assigned to cluster
K
✓ µ𝒌 : The centroid (mean vector) of cluster K.

✓ 𝒙𝒊 − µ𝒌 𝟐 : The squared Euclidean distance

(x - cluster center).

Santiago Londoño López

Politecnico di Torino
DISEG – ArtIStE laboratory
 Response Quantities study for the Maximum Rotation-
Independent Principal Ground Acceleration:
The study of a NLTHA of (MDOF) Reinforced Concrete 3D model to examine the correlation
between standard / raw inputs and The Maximum Rotation-independent principal ground
acceleration.

• 3 Reinforced Concrete Structures

• Case 1: Regular – Rectangular Shape - 8 Storeys
• Case 2: Regular – Square Shape - 8 Storeys
• Case 3: Irregular – L Shape – 8 Storeys.

• Input: >200 Accelerometer pairs Magnitude > 4.5

4 orthogonal pair GM
• Xstru-Ystru oriented as XGM_princ-YGM_princ
• Xstru-Ystru oriented as XGM_recorded-YGM_recorded
• Xstru-Ystru oriented as XGM_princ-YGM_princ + 𝛳 =30°
• Xstru-Ystru oriented as XGM_princ-YGM_princ + 𝛳 =60°
Santiago Londoño López
Politecnico di Torino
14
DISEG – ArtIStE laboratory
 Preliminary results:
Test Structure: 8-Storey Rectangular Shape - Regular – Ground Motion #18
Shear Vx at the base for Corner Column01 Shear Vy at the base for Corner Column01
4 Considered Orientations 4 Considered Orientations

Santiago Londoño López

Politecnico di Torino
DISEG – ArtIStE laboratory
 Preliminary results:
Test Structure: 8-Storey Rectangular Shape - Regular – Ground Motion #18
Inter-story Drift – X Direction for Inter-story Drift – Y Direction for
4 Considered Orientations 4 Considered Orientations

Santiago Londoño López

Politecnico di Torino
DISEG – ArtIStE laboratory
 Case of Study 3
Optimization Of Steel
Exoskeletons For The Seismic
Retrofit Of Reinforced Concrete
Structures Via Genetic
Programming
(Olivo, et al, 2024) - ArtIStE
 EXOSKELETONS
avoid the interruption of the applied from outside
activities carried out inside steel additive structures
of the building
low construction time
suitable for a holistic and cost
intervention: structural,
energetic, architectonic inspired by nature:
biomimicry

ORTHOGONAL PARALLEL

Santiago Londoño López

Politecnico di Torino
DISEG – ArtIStE laboratory
 OPTIMIZATION TOOL
OBJECTIVE
FUNCTION

DESIGN
VARIABLES

PENALTIES

Santiago Londoño López

Politecnico di Torino
DISEG – ArtIStE laboratory.
OPTIMIZATION ALGORITHM & FEM ANALYSES

Santiago Londoño López

Politecnico di Torino
DISEG – ArtIStE laboratory
RESULTS: INTER-STOREY DRIFT

Santiago Londoño López

Politecnico di Torino
DISEG – ArtIStE laboratory
RESULTS: BASE-SHEAR

Santiago Londoño López

Politecnico di Torino
DISEG – ArtIStE laboratory
 Case of Study 4
Structural Damage Detection In
Bridges Under Operational And
Environmental Variability By
Means of Machine Learning
Identification

(Zunino, et al, 2025) - RealSmart

 Context
Traditional methods: New frontier:

• Manual inspection • AI & Machine

SHM

• Modal Analysis Learning for

• Thresholding automatic damage
detection
Objective
Develop a robust unsupervised SHM
framework to detect damage using only real
acceleration data collected under traffic and
temperature effects.

Santiago Londoño López

Politecnico di Torino
DISEG – ArtIStE laboratory
Case Study

The Yonghe Bridge is a cable-stayed bridge

equipped with a long-term structural health
monitoring system. Acceleration data were
collected both before and after a verified
damage event, enabling a direct comparison
of the bridge’s dynamic response under healthy
and damaged conditions.

Santiago Londoño López

Politecnico di Torino
DISEG – ArtIStE laboratory
Methodology ▪ Input: Raw accelerations under EOV
▪ Processing:
Variational Mode Decomposition (VMD)
Hilbert-Huang Transform (HHT)
Principal Component Analysis (PCA)
Clustering with K-means
▪ Output: Damage Detection

Santiago Londoño López

Politecnico di Torino
DISEG – ArtIStE laboratory
 Variational Mode
Decomposition
▪ Decomposes signal into band-limited
IMFs.
▪ Robust to noise, avoids mode mixing.
▪ Requires an optimization strategy for
setting parameters.

Santiago Londoño López

Politecnico di Torino
DISEG – ArtIStE laboratory
HILBERT-HUANG ▪ Extracts instantaneous frequency and amplitude from
each IMF, used as damage-sensitive features.
TRANSFORM ▪ Provides a time–frequency representation ideal for
non-stationary signals.

Santiago Londoño López

Politecnico di Torino
DISEG – ArtIStE laboratory
 Principal
Components
Analysis
▪ Removes environmental and
operational effects, which
dominate the variance.
▪ Discarding the first principal
component (PC1), we isolate the
components more likely related to
structural behavior.

Santiago Londoño López

Politecnico di Torino
DISEG – ArtIStE laboratory
Clustering: K-Means

Undamaged Condition Damaged Condition

Santiago Londoño López

Politecnico di Torino
DISEG – ArtIStE laboratory
Clustering: K-Means

DC values Windowing

Box Plot
Windowing

Damage Index
Santiago Londoño López
Politecnico di Torino
DISEG – ArtIStE laboratory
Thanks for your
Attention
Questions?
Speaker:
Santiago Londoño López
MSc S.E., C.E.

[Link]@[Link]