Abschlussarbei­ten

Linda Kevine Guiameugne Guabiapsie

Dr. Roman Rezaev (Institute of Construction Materials)

Develop a Machine Learning and Deep Learning framework for estimating the yield stress and plastic viscosity of fresh concrete using information extracted from image processing of flow table test video recordings and reference rheological measurements obtained from a rotational rheometer.

Rheometers provide measurements of yield stress and plastic viscosity but are expensive and impractical for routine field use. Although the flow table test is simple and widely used, it provides only indirect information about rheology, and conventional empirical correlations are often insufficient, particularly for plastic viscosity, creating the need for more robust prediction approaches.

Joachim Wisnewski, M.Sc. (Institute of Steel and Timber Construction)
Dipl.-Ing. Lars Feulner (Züblin Stahlbau GmbH)
Dipl.-Ing. Arend Kiel (Züblin Stahlbau GmbH)

•  Develop an automated framework
•  Generate a parametric model in RFEM 6
•  Check its structural behavior
•  NSGA-II optimizes its self-weight 

The implementation of a multi-objective optimization framework using NSGA-II through Pymoo successfully explored the design space of ‘Thick-plate bridge’. During the optimization process, ten Design Variables were optimized in order to minimize the bridge’s self-weight in terms of steel and to maximize its structural performance. It was also found that implementing the second Objective Function and its respective Penalty Function improves the speed of convergence, reducing it almost to half the time. The implementation of eleven Functional Constraints (seven from the utilization ratio of the structural components and 4 from geometrical relationships) also involves the evolution process, giving NSGA-II feedback at the end of every generation. Moreover, the Functional Constraints played an important role in the final result because they ensured that it would pass the internal stability (buckling) in the post-processing phase.

Because stability verification takes more than 10 times strength, it was shifted to post-processing. NSGA-II was configurated to ensures the first of the sorted feasible solution become the final. Pareto fronts demonstrates stability, algorithm effectiveness and a well distribution of solutions.

Darwin German Caicedo Perez

Marcelo Pintado, M.Sc. (Scale Company)
Dipl.-Ing. Marko Thiele (Scale Company)

To develop a method to automatically detect and classify geometric changes in automotive finite element models, whose output will be used to construct prompts for large language models in order to generate detailed and standardized documentation entries.

During the vehicle design phase, finite element models are modified to improve their performance in specific load cases, such as crash tests. Documentation meant to describe these changes is often overlooked due to negligence, time constraints, or low prioritization. As a result, valuable insights from previous simulations are lost, disrupting knowledge transfer between past and future engineers and leading to missed opportunities to replicate successful outcomes.

This work presents a method to automatically detect and classify geometric changes in digital vehicle models. Regions with changes are converted into point clouds and analyzed using machine learning models based on the PointNet and PointNeXt architectures to identify the type of geometric modification from a predefined set of change classes. The classification results are combined with the simulation engineer’s initial documentation entry and structured into a prompt which will be processed by a large language model to generate standardized and detailed documentation. The geometric machine learning models are trained on a dataset created by SCALE GmbH, based on iterative optimizations of a Toyota Yaris sedan model, where pairs of meshes (base and modified) of the parts of the vehicle are classified into bead, hole, and sheet change categories.

The proposed method can detect and classify geometric changes in automotive models, categorizing modifications into beads, holes, and sheet subclasses. Among the tested input and architecture configurations, PointNet achieved a maximum classification accuracy of 61.21%, while PointNeXt reached 52.99%.

Linda Kevine Guiameugne Guabiapsie

Develop and validate an automated, image-based approach to analyse the dynamic flow behaviour of fresh concrete during the flow table test and establish correlations with rheological parameters measured by a rotational rheometer. Thus, transforming the flow table into a practical surrogate for rheological assessment.

Rheometers provide measurements of concrete’s yield stress and plastic viscosity but are expensive, heavy, and unsuitable for on-site use. In contrast, the flow table test is simple and widely used to assess workability, yet it offers only indirect correlations to rheological properties. While larger spreads generally indicate lower yield stress, viscosity cannot be reliably captured unless the time evolution of spreading is analysed. Moreover, concretes with different flow behaviours can exhibit the same final spread, highlighting the need to incorporate time-dependent analysis for more accurate rheological correlation.

The flow table test was recorded to capture the dynamic flow behaviour of fresh concrete. Frames were extracted from the videos to balance data size and temporal accuracy. Image coordinates were corrected to align with the horizontal plane for consistent diameter measurement. Segmentation was then applied to isolate the concrete for precise analysis, and image-based measurements (concrete´s diameter) were correlated with rheometer data to link visual observations to rheological properties.

Image-based and laboratory measurements of diameters show good agreement, with minor deviations (–10% to +10%) primarily due to table corner detection and camera setup. The image-based method tends to slightly overestimate the diameter in most cases. Correlation between the concrete’s diameter and its rheological properties (yield stress and plastic viscosity) was non-conclusive. Therefore, a dataset was prepared for training an Artificial Neural Network (ANN).

Mukhammad-Bobur Zokhidov

Designing truss bridges involves balancing multiple conflicting objectives, such as ensuring cost efficiency while improving the strength of the structure.  Systematic Multi-objective Optimization Approaches (MOAs) are needed to generate efficient, feasible, and well-balanced designs under real-world constraints.

Traditional truss bridge design methods often address objectives separately, prioritizing either cost, strength, or serviceability in isolation. In real-world projects, the design process is limited to a narrow range of predefined bridge types, restricting flexibility in determining the optimal shape. Without a systematic optimization strategy, the resulting designs may be either over-engineered, leading to unnecessary costs, or underperforming, compromising safety and durability. A robust multi-objective framework is therefore required to effectively balance these competing demands.

Two advanced algorithms, NSGA-II and MOPSO, were studied and additionally three MOAs were implemented using the Python-based Pymoo and Platypus frameworks, with OpenSeesPy performing structural analysis. Constraints applied included the kinematic stability of the truss structure, allowable tensile and compressive stresses, and maximum permissible deflection. The MOAs were first tested on benchmark truss problems to evaluate performance. The proposed approach was then applied to Ruhr Canal Bridge using different truss bridge types to determine optimal member sizing and geometry that achieve the best trade-off between minimum structural volume and minimum deformation, while satisfying strength and serviceability constraints. The bridge model was developed as a single-span steel truss system with a concrete slab deck and simplified load case was calculated according to Eurocode.

Performance measures for different algorithms using  truss benchmark problem.

Darwin German Caicedo Perez

Akhil Pillai, M.Sc. (Scale Company)
Dipl.-Ing. Marko Thiele (Scale Company)

To generate a structured dataset consisting of simulation inputs and outputs, along with semantic information, for the fine-tuning of a large language model (LLM) that will support the standardization of the documentation process in automobile development.

The large number of simulations performed during the design phase of vehicles often results in documentation that is unclear and lacks detailed description of the modifications made to the FEM models of vehicles. These changes, intended to improve vehicle performance in specific load cases like crash tests, are sometimes overlooked because documenting them is time-consuming and not always given priority. Consequently, important insights from previous simulations are lost, which disrupts the knowledge transfer between past and future engineers, causing repeated work and missed chances to reproduce successful outcomes.

This project focuses on generating a dataset containing simulation inputs that capture the modifications made to FEM models, along with corresponding outputs for visualizing their results. Additionally, relevant semantic information is added: the reasoning behind each change, the expected outcomes and comparisons between the performance of the base and current versions. To create this dataset, a development process is replicated using a Toyota Yaris FEM model of public domain.

Numerous modifications were made to the structural components of the initial vehicle model to enhance its performance in the IIHS side crash test and the USNCAP front crash test. Performance evaluation was based on various metrics: for the side crash test, these included intrusions, stress and strain field visualizations, and structural element failure in the impact area; for the front crash test, the indicators included accelerations across different car regions and firewall intrusions. These metrics were used to assess improvements between the base and modified models. Additionally, an initial workflow was developed for constructing prompts that will serve as the input for a future fine-tuned LLM, which will generate enhanced documentation entries. The dataset generated in this project will serve to fine-tune that LLM.

Ivan Burtovoi

Dipl.-Ing. Tobias Mansperger (LAP-Consult Beratende Ingenieure VBI AG)
Dipl.-Ing. (FH) Martin Hohmann, M. Sc. (LAP-Consult Beratende Ingenieure VBI AG)
Petr Novak, M. Sc. (LAP-Consult Beratende Ingenieure VBI AG)

Anargul Jamankulova

Multi-Objective Evolutionary Algorithms (MOEAs) is a potent approach in civil engineering optimization, addressing complex problems with multiple conflicting objectives. Inspired by the theory of natural evolution, MOEAs efficiently find solutions that balance competing goals, such as cost-effective and eco-friendly structures. Selecting the algorithm to optimize the structure primarily focuses on effectively managing various real-world challenges and ensuring the algorithms' performance meets desired criteria.

• Investigate new advanced methods for algorithmic operations to improve performance and address diverse challenges in MOPs
• Study critical factors influencing algorithm efficiency and effectiveness in solving various optimization problems
• Optimize these factors according to real-world MOPs

• Performance of algorithms for different test problems
• Optimal algorithm parameters according to performance measures

Ivan Burtovoi

Dipl.-Ing. Tobias Mansperger (LAP-Consult Beratende Ingenieure VBI AG)
Dipl.-Ing. (FH) Martin Hohmann, M. Sc. (LAP-Consult Beratende Ingenieure VBI AG)
Ing. Petr Novak (LAP-Consult Beratende Ingenieure VBI AG)

To create a practical automated tool for optimizing the weight of steel in a composite bridge during the preliminary design stage

The project focused more on estimating material consumption than exhaustive calculations for every bridge element, balancing computational time and accuracy. It aimed to explore optimization techniques for steel components in composite bridges, considering material properties, design methods, and construction techniques. The research also addressed challenges and opportunities in bridge construction posed by environmental and economic factors.

The optimization process introduces a Python-based tool that seamlessly integrates with SOFiSTiK software, augmenting its capabilities. Leveraging Python's versatility and SOFiSTiK's computational power, critical data such as internal forces and stress values are efficiently extracted for rigorous stability assessment of bridge elements. A key innovation lies in the adoption of a simplified model during the tender phase, ensuring swift design iterations without compromising analysis integrity, indicating an already developed solution.

This project focuses on the optimization of a composite road bridge, as specified by the engineering firm LAP-Consult. The bridge in question uniquely integrates a concrete deck, specified as C45/55, with a steel box girder composed of S355 grade steel. This particular combination was selected during the tender phase of the project, primarily to estimate material consumption, and does not represent a detailed design of the bridge. The bridge spans a total length of 939 meters, comprising of nine segments with lengths of 67, 100, 110, 125, 135, 125, 110, 100, and 67 meters, respectively. The structural design incorporates two independent continuous beams connected to the piles through pinned connections, forming the bridge’s static system. The final analysis, encompassing multiple simulation cases, demonstrated a substantial reduction in weight, resulting in an impressive improvement of 135 tons compared to the initial weight across various scenarios. This reduction signifies significant cost savings, indicating a financially viable outcome. Further stability checks are advised for the subsequent stage of the project.

Dr.-Ing. Birgit Beckmann (Institute of Concrete Structures)

Application and evaluation of suitable approaches to analyze the strain time series and to cope with the noise and artifacts

Dynamic load tests, like drop tower experiments, are commonly used to investigate the structural response and material behaviour of concrete. These experiments typically produce time-dependent data in the form of time series. However, the resulting time series often contain noise and artefacts that can reduce the usefulness of the gathered information and obscure the actual underlying strain behaviours of concrete elements.

Firstly, the analytical time series was generated to assess the effectiveness of various techniques for time series analysis. Various methods, including polynomial regression, spline regression, moving average models, exponential smoothing, and fast Fourier transform (FFT), were utilized to detect and adjust any present instances of noise and artefacts. After evaluating different models based on the comparison of their error metrics, the most accurate approaches for the analytical dataset were selected and applied on the strain time series obtained from a drop tower experiment with concrete slabs.

The assessment results indicated that moving average models were the more precise approaches for the analytical dataset, although they could not entirely eliminate all noise and cyclic components in the time series.

Oracides Felício Adriano, B. Sc. (Oracides Adriano Engenharia Especial Ltd)

• Use of a Large Eddy Simulation of the mountainous terrain to perform the dynamic analysis of the bridge
• Investigate aerodynamic derivatives and admittances under varying angles of attack and reduced wind speeds. tudy the bridge's stability and response to stochastic wind for six wind directions
• Evaluate the bridge's preliminary design to assess its resistance to the dynamic wind loading

Studying the bridge's stability and dynamic response to wind is of immense importance in bridge engineering. The study of motion and flow-induced loads has been primarily studied in wind tunnels, which was not the case for this bridge. Therefore, this master thesis aimed at developing tools to assess the motion and flow-induced loads using the software OpenFOAM. Additionally, a tool was developed to evaluate the bridge's stability and total dynamic response for different wind characteristics in different wind directions.

The implemented solution can be summarized in the following steps.

• Eigenvalue analysis of the bridge under the mean wind load, one eigenvalue analysis for each wind direction
• Evaluation of the aerodynamic derivatives for different angles of attack and reduced wind speeds
• Evaluation of aerodynamic admittance functions for different angles of attack considering a turbulent inlet
• Assessment of the bridge’s dynamic and static stability for an increasing wind field considering different eigenvalues and wind characteristics for six wind directions
• Assessment of the dynamic response and design of the bridge for the six different wind directions

• Motion and Flow-Induced Forces: The calculation of the aerodynamic admittance functions did not match the expected behavior for a bridge’s cross-section.
• Stability And Response Spectrum Analysis: The developed tool evaluates the bridge's stability and calculates the maximum displacement it gives the impact of the aerodynamic damping and stiffness on the total damping and stiffness of the system (air + structure). The results showed that, under two wind directions, the bridge loses vertical stiffness due to the mean wind, which causes a reduction in the total damping of the system due to the motion-induced portion of the load.

Dipl.-Ing. Marko Thiele (Scale Company)
Akhil Pillai, M. Sc. (Scale Company)

Improve the synergy between artificial intelligence and the automotive construction industry by developing a machine learning model that can predict similar parts and allow updates from continuously growing data

Machine learning algorithms to recognize geometric parts are not yet implemented in the productive sector. These models have the potential to improve quality, performance and efficiency in the automotive industry, by reducing the risk of errors, improving quality control and streamlining the manufacturing process.

A machine learning model that includes augmentation data with the iterative closest point (ICP) as geometry alignment, classifier customization of the PointNeXt architecture, innovative training with non-supervised clustering and smooth labeling and transfer learning as model generalization.

Training dataset with 3 automobiles

Dr.-Ing. Markus Uhlig (Institute of Geotechnical Engineering)

• TensorFlow models consume more computational time but have better accuracy than step by step built up models.
• In the parametric study, the following conclusion can be made:
  01. Number of hidden layers has a slight effect on the performance, and the effect of the percentage of the testing data is almost neglectable
  02. The larger batch size, the faster the model converges and more computational time is required
  03. The higher learning rate, the faster the model converges and less computational time is required
  04. Using Max-Min normalization give better performance than standard normalization
  05. Using segment size of 25, batch size 10 and learning rate 0.001 gives the best performance

Oracides Felício Adriano, B. Sc. (Oracides Adriano Engenharia Especial Ltd)

• Study the aerodynamic properties of a mountainous region required for the future design of a footbridge
• Application of the aerodynamic properties into virtual wind tunnel section models
• Study the bridge's response to different cross-sections

In bridge engineering, the study of the construction site's wind properties and the bridge's susceptibility to wind and motion-induced loads is of immense importance. The study of wind properties and motion-induced loads has been primarily studied in wind tunnels. However, full-scale wind tunnel models and section tests might make the project impracticable for small projects like pedestrian footbridges. Therefore, this project aimed to develop tools to make those studies possible in an open-source CFD software, OpenFOAM.

The implemented solution can be summarized in the following steps.

• Development of a turbulent inlet boundary condition that accurately represents the general wind characteristics on site
• Simulation of the wind on the terrain in several directions, using the developed boundary condition
• Evaluation of wind properties on several points along the bridge span
• Use the wind properties to study different bridges' cross-sections
• Calculate force coefficients for bridges' cross-sections
• Study of aerostatic instabilities for different bridges’ cross-sections

All wind properties (e.g., mean wind speed, turbulence intensities, mean angle of incidence, and all wind spectrums) and force coefficients were analyzed and preserved for future dynamic analysis. The wind spectrum showed a significant decrease in the high-frequency range, which was corrected by considering that the high-frequency range has very little influence on the final calculation of the standard deviation of the instantaneous wind speed.

Anvar Mohamed Aslam Sha

Prof. Dr.-Ing. habil. Uwe Reuter (Faculty of Civil Engineering, IT Service Center)

Prof. Dr.-Ing. habil. Wolfgang Weber (Helmut Schmidt University Hamburg)

Dipl.-Ing. Marko Thiele (Scale Company)
Akhil Pillai, M. Sc. (Scale Company)

Evaluate two machine learning algorithms using point cloud analysis identifying similar parts from 3D automobile CAD data to improve the vehicle design process

New technologies and equipment are developed every day in the automotive industry, leading to an increasing number of models and parts manufactured, which are similar due to an automobile’s topography. Waste prevails in this production, where due to the lack of time and resources, raw materials are used to manufacture new parts instead of taking advantage of the similarities and recycling.

Data starts with a 3D mesh converted into a 3D point cloud and is further augmented to train the neural network. This work uses PointMLP and 3DmFV approach. This last one requires a Fisher vector representation.

Influence of sample and part numbers and the architecture capacity for each machine learning algorithm

Reducing the effort of geometry conversion from CAD into an applicable mesh for FEA using  isogeometric analysis (IGA) technique representing a great connection between Computer-Aided Design and Finite Element Analysis (FEA)

Implementation of the isogeometric analysis using the basis functions of the same ones that are used for geometry representation this counts as a standout point in comparison with FEM along with the improved representation that is done by the Non-Uniform Rational B-Splines (NURBS) basis function for the computational space. The basis function NURBS require a third solution space in order to assist the numerical integration done in parent space another space so called parametric space is introduced and mappings therefore required to obtation stiffness matrix and force vector.

Simplifying the interaction with computer aided geometric design (CAD), NURBS basis functions were studied and implemented compuationally as they offer continuity control for basis functions, in order to to dicuss solutions effeciency  the method of refinements that are available in Isogeometric analysis were represented through a computational implementation based on Python language, the analysis method of IGA was performed and a certain linear elastic problem for a shape with curved section was evaluated for solution fields such as displacement and stress.

The implementation of an example using isogeometric analysis illustrated the exact solution feature of the method, the choice of a problem with existing analystical solutions was taken as a benchmark for the resulting values of stress in all directions, the results showed very accurate solutions with few refinements techniques, howevert the coarse mesh with no further refinements approaches acceptable results that are very close to analytical solutions.

Prof. Dr.-Ing. habil. Uwe Reuter (Fakultätsrechenzentrum der Fakultät Bauingenieurwesen)
Prof. Dr.-Ing. habil. Frohmut Wellner (Institut für Stadtbauwesen und Straßenbau)
Gustavo Canon Falla, M. Sc (Institut für Stadtbauwesen und Straßenbau)
Dipl.-Ing. Paul Bolz (Institut für Stadtbauwesen und Straßenbau)

Reduzierung des für die Bestimmung mechanischer Asphalteigenschaften benötigten Zeitaufwands

Das mechanische Materialverhalten des Stoffgemischs Asphalt ist aufgrund seiner Heterogenität überaus komplex. Die einzelnen Bestandteile weisen ein sehr unterschiedliches Materialverhalten auf. Es liegt eine ausgeprägte Temperatur- und Belastungsfrequenzabhängigkeit vor. Weiterhin hat die zufällige Anordnung und Verteilung der Gesteinskörner und Luftporen eine gewisse Unsicherheit zur Folge. Die Bestimmung der mechanischen Eigenschaften von Asphalt in Laborversuchen sowie in physikalisch-numerischen Modellierungen ist deshalb mit einem erheblichen Zeitaufwand verbunden.

Zunächst wurden der absolute E-Modul |𝐸|und der Phasenwinkel Φ für Asphaltgemische mit variierenden Materialzusammensetzungen und in Abhängigkeit der Belastungsfrequenz sowie der Temperatur in stochastischen Finite-Elemente-Simulationen bestimmt. Im Anschluss erfolgte eine Metamodellierung unter Anwendung künstlicher neuronaler Netze (Convolutional-Neural-Networks, CNNs) auf die Simulationsergebnisse. Die stochastischen Eigenschaften der Modellierungsaufgabe wurden durch zwei Metamodellvarianten in unterschiedlicher Weise berücksichtigt.

Durch die Metamodellierung konnte eine Reduzierung des Zeitaufwands gegenüber den Finite-Elemente-Simulationen unter gleichzeitiger Einhaltung einer hohen Vorhersagegenauigkeit erreicht werden.

Finding multistability in tensegrity structures through form-finding and parameter-finding

Choosing the most suitable approaches for multistable tensegrity structures requires study of literature related to tensegrity structures, design optimization, and numerical methods for form-finding. The suitability is proven by implementing the approaches, including optimization of parameters characterizing specific properties of the structure and the demonstration of the implemented methods on example of practical use. Visualization of the obtained results is also needed to provide a better understanding.

Both genetic algorithm and stiffness matrix form-finding (SMFF) methods are implemented in the multistable form- and parameter-finding. The genetic algorithm generates random numbers that act as variable inputs for the SMFF method in nodal coordinates. The coordinates will be refined through non-linear equation resulting in new nodal coordinates of a shape in equilibrium state and stable condition. The obtained result will be evaluated by the fitness functions, which are maximization of volume  and minimization of nodal displacements for form- and parameter-finding algorithms, respectively.

• The multistable set of tensegrity shapes is successfully obtained throughout the form-finding process.  The diversity of the solutions are shown in the graph along the generations.
• The parameter-finding results are showing that the suitable cable parameters for multistable tensegrity structures are obtained, while new stable configurations with diffe-rent parameters are also found.

Dr.-Ing. Christian Jenkel (FRILO Software GmbH)
Dipl.-Ing. Matthias Friedrich (FRILO Software GmbH)

• Study and choice of machine learning approaches which are suitable for problems of pattern recognition
• Implementation and evaluation of machine learning models with special consideration of effectivity and efficiency
• Application on examples of practical use and satellite data

Tracking of factories scattered all over the world is a useful instrument for the management of processes by the factory owners. The analysis of satellite data is one possibility to track factories. The evaluation of satellite data succeeds by methods of pattern recognition. Common approaches of pattern recognition are based on methods of machine learning.

Neural network is effective machine leaning method for pattern recognition. The fundamental algorithms of neural networks are Feed Forward, Gradient Descent and Back Propagation algorithms.

Artificial Neural Network (ANN) and Convolutional Neural Network (CNN) on GTSRB multi-category classification problem

Prof. Dr.-Ing. habil. Uwe Reuter (Fakultätsrechenzentrum der Fakultät Bauingenieurwesen)
Prof. Dr.-Ing. habil. Frohmut Wellner (Institut für Stadtbauwesen und Straßenbau)
Gustavo Canon Falla, M. Sc (Institut für Stadtbauwesen und Straßenbau)
Dipl.-Ing. Paul Bolz (Institut für Stadtbauwesen und Straßenbau)

Erstellung eines Metamodells für die Prognose von Asphalteigenschaften

Die Vorhersage von Asphalteigenschaften ist sehr komplex. Asphalt ist ein heterogenes Stoffgemisch und seine Materialeigenschaften sind von einer Vielzahl von Eingangsgrößen abhängig. Laborversuche, mit denen der Einfluss der Eingangsgrößen auf die Materialeigenschaften bestimmt werden kann, sind zeitlich sehr aufwendig. Daher existieren kaum Daten, die für die Prognose der Asphalteigenschaften herangezogen werden könnten.

Durch die numerische Modellierung von Asphalt als heterogenes Stoffgemisch kann zunächst eine Datengrundlage geschaffen werden. Unter Anwendung von Methoden des maschinellen Lernens wird im nächsten Schritt ein Metamodell zur Prognose der Asphalteigenschaften in Abhängigkeit der Eingangsgrößen erstellt. 

Die KNN-Modelle konnten bezüglich der Genauigkeit bessere Ergebnisse als die SVR-Modelle liefern. Jedoch besteht für die SVR-Modelle noch Optimierungspotential hinsichtlich der Modellparameter. Um ein brauchbares Metamodell zu erstellen, ist außerdem die Generierung weiterer Daten notwendig. Die Laufzeit der SVR-Modelle war im Vergleich zu den KNN-Modellen wesentlich kürzer (0,04 Sekunden vs. 15 Minuten).

Implementation and evaluation of multi-objective decision support approaches in environmental engineering using the lake problem

Environmental engineering has many sectors where decision support approaches could be more applied. Also, the human actions attempting to increase their economic benefits create severe environmental problems currently appearing in many forms, which must usually balance uncertainties, tipping points and conflicting thresholds. Therefore, better management policies regarding ecological consequences and economic benefits should be the goal of this century.

In this research, a decision problem is resolved by four multi-objective robust decision support approaches which are variations of robust decision making. All analyses are implemented in Fortran language and different aspects will compare the methods as the time of analysis, robustness and quality results. Moreover, a sensitivity analysis is performed to a better understanding of the uncertainties of the problem. A commonly used environmental decision problem, known as Lake Problem, will be evaluated by the approaches.

Optimized structural design of wind turbine blade using applicable and cost efficient approaches

To make an optimization process applicable to real-life situations, more than one design objective must be considered. In this master thesis design objectives are: buckling strength, flapwise deflection, total cost and mass of the blade. The varying parameters are thickness and fiber orientation of composite layers and thickness of core material. For finding the optimized design the non-dominated sorting genetic algorithm II (NSGA-II), the multiple objective particle swarm optimization (MOPSO), and the single objective ε−constrained method were used.

The selected blade is an open source model of the DTU 10 MW reference turbine provided by the Danish Technical University. The optimization methods were applied for the entire radial length of 89 m. To obtain the values for the objectives, three finite element software were used: ABAQUS for buckling analysis, BECAS for strength analysis with mass and cost estimation, and FRANS for stiffness analysis. The preprocessor for 2D cross section analysis is a mesh generator Shellexpander.

Dipl.-Ing. Marko Thiele, SCALE GmbH
Jana Büttner, M.Sc., Porsche AG

Reduce the computational time of large FEM models while maintaining a desired accuracy

The automobile industry is highly dependent on numerical modeling. Companies perform numerical simulations for crash analysis in the preliminary phases of design and processing, because it eliminates the need to create and crash expensive physical prototypes. Crash analysis is even performed several times using different scenarios, all of which requiring large computational times. Optimization, which is a time consuming process by itself, is also often performed during the different stages of a project. This further adds to the scale of computational time in the automobile industry. The longer it takes to perform such numerical processes, the larger the time-to-market of a certain vehicle model and eventually, the higher the cost.

In order to speed up the calculation and assessment process, Model Order Reduction (MOR) techniques are utilized. MOR reduces the complexity of a system while maintaining its input-output behaviour to a certain accuracy. It aims to find a reduced order model (ROM) in a reduced subspace ℝ𝑘R^k from the full order model (FOM) by the aid of a transformation T. In structural dynamics, the reduced vector of displacements 𝒙𝑘(𝑡)▁x_k (t) where 𝑥𝑡=𝐓𝒙𝑘(𝑡) 〖▁x (t)=T▁x〗_k (t) yields the reduced equation of motion 𝐌𝑘𝑥𝑘(𝑡)+𝐂𝑘𝑥𝑘(𝑡)+𝐊𝑘𝑥𝑘(𝑡)=𝑓𝑘𝑡M_k ▁(x ̈ )_k (t)+〖C_k ▁(x ̇ )〗_k (t)+K_k ▁x_k (t)=▁f_k (t) where the system matrices are obtained □𝑘=𝐓𝑇□𝐓□_k=T^T□T.
There exist different MOR methods which handle linear systems and non-linear (with regards to geometry, material, and boundary) ones. Methods handling linearity are much more mature and have been in service for a long time. Although, a crash exhibits several non-linear parts, there exist large portions of the vehicle that behave linearly, thus are candidates for linear MOR. The finite element software LS-DYNA is often used in crash analysis. It incorporates two linear reduction approaches, Superelement (SE) and Linearized Flexible Body (LFB). SE can be based on static condensation or Craig-Bampton reductions while LFB can be based on modal truncation or Craig-Bampton, empowered by rigid-body-motion displacement. The methods were first tested on a structural frame impacted by a sphere.

• SE performs better in small displacement prob-lems
• SE fails in large displace-ment problems
• Interactive nature of reduced part (case 3) improves accuracy
• The  number of eigen values influences LFB reduction more than SE

Probabilistic methods will be used to analyze if reduced permanent action partial safety factor, PSF, values may be used for larger structural elements typically found at power plants. Additionally, the sensitivity of the models to changes in input parameters and solution methods will be explored.

In Germany power plants are designed based on VGB guideline R 602 U. This standard follows the format of the Eurocodes, with some adjustments to account for special conditions of power plant construction, for example the difference in applied actions/loads. The PSFs were taken directly from the Eurocodes (EN DIN 1991-1-1). Given the larger sizes in structural elements in power plants, it follows that the construction deviations will be smaller when taken as a percentage of the member size when compared to general structures. Therefore, a possibility in the reduction of the permanent action PSF for power plants and other industrial buildings will be evaluated and the bias from different approximation methods will be explored.

The first order reliability method, FORM, creates the basis for the analysis because it yields the sensitivity factors (αα) needed to calculate PSFs (𝛾γ). It also estimates the reliability index (ββ). Additional methods are for comparison, since FORM results are approximate. These methods are the second order reliability method (SORM), the Monte Carlo method and Gaussian Quadrature. In order to use FORM/SORM on an implicit limit state function, two different types of regression functions were used. The first type was linear regression with a quadratic polynomial function and the other was support vector regression, SVR.

The FORM-based comparison of the two model slabs was carried out with the variation of different components of the model including varying the target variable of the regression and varying other related variables (such as live load). The results and impact on the calculated PSFs are shown below. The horizontal lines indicate PSFs calculated based on assumed sensitivity factors from DIN EN 1990-1-1, and represent conservative upper limits. The results show slightly smaller values of the permanent action PSF, in average, for the power plant model.

Kaike Monteiro

Implementation and evaluation of decision support approaches using the example of bridge construction and maintenance

In this research, bridges will be analyzed by one method of 4 types of decision support approaches. Therefore, the bridge will be evaluated by a reliability analysis (exact probabilistic – level 3 reliability), a cost analysis (cost-effectiveness analysis), a risk analysis (event-tree analysis), and a sensitivity analysis (variance-based sensitivity analysis), in this order. The utilization of four methods from different decision support approaches will show the amount and variety of information that a decision-maker can obtain to make the optimal decision possible with an efficiently implementation.

Kamila Tlegenova

Implementation of benchmark example for obtaining the minimum mass of the blade using structural optimization approaches

A sustainable design of wind turbines requires, amongst others, an optimization of the aerodynamic, the structural, and the aero-servo-elastic design. In order to find the optimum design in a given time, efficient approaches for optimization of the design parameters are needed, which require the definition of corresponding objective functions. The aim of this project work is to analyse the different aspects of design optimization of wind turbines and to implement selected objective functions, which can be used as benchmark examples for optimization approaches.

The variable selection for this problem requires certain methods like Sensitivity Analysis. In this project the thicknesses of different materials that construct structural regions of the blade cross section such as caps, webs and etc., were selected as random variables. After, the reduced Monte Carlo optimization can be performed and compared with the full Monte Carlo optimization. The strength and mass of the blade are calculated using MATLAB functions by BECAS which uses ABAQUS 3D shell model and 2D grid created by Shellexpander python script.

After performing Sensitivity Analysis only five variables were selected for the reduced Monte Carlo Optimization. However, according to time cost, the difference is not useful and rather reduced MC takes more time because of Sensitivity Analysis.

Minimization of structural weight of trusses by considering displacement, stress and kinematic stability constraints

To minimize the mass of the steel structure it is necessary to deal with the uncertainties regarding size, shape and topology of the truss with consideration of deflections and feasibility of structure.
For member load bearing capacity verification European standard EN 1993-1-1 is used considering axial and flexural criterions.
Frilo RSX software was used for FEM calculations. Algorithm was written in C# language.

The alogrithm used was specially designed for truss optimization. On first stage shape and topology is optimized on second stage-size.

The approach used for algorithm proved to be capable of optimizing trusses if tunned properly. The usage of discrete cross-sections make optimization process faster but it was also evident that in such cases there is still possibility to minize mass further if cross-sections are allowed to be redesigned by changing dimensions but keeping for example: cross-sectional area as constant.

Zeidoun El Khatib

Improve the efficiency and accuracy of a finite-element-based optimization program

The Finite Element Method (FEM) is used daily in different fields and applications. However, applying it shall not be done blindly since the results are subject to error. Not only is FEM useful in analyzing structures, but combined with optimization strategies, it serves as an excellent tool to conceptualize and optimize structures in order to take the best out of them while saving material and resources. With the poor management of resources in the world, structural optimization has become of great importance. Yet, optimization is a time demanding numerical iterative process, thus efficiency is of great concern. A given optimization program suffers from problem size restrictions due to memory allocation and some accuracy problems.

To overcome the memory allocation problems, several measures were implemented. The timely de-allocation of dynamic variables ensured the clearance of the heap memory and prevented memory leaks that result when such arrays aren‘t de-allocated. Global variables were used to overcome variable redundancy induced by subroutines. Also, input and output into files were controlled and minimized. The reduction of the system due to boundary conditons was boosted by the aid of logical variables. Last but not least, since the stiffness matrix in FEM problems is sparse, the memory requirements were reduced by storing only non-zero stiffness values. A data structure, Compressed Sparse Row (CSR), was implemented for this purpose; its formation was done directly and indirectly in order to validate the code and reflect on memory savings. Solving the system was done by UMFPACK, a sparse direct solver.
To trace the accuracy problems, the influence of the system solution method was also studied. In addition, care was given to singluarity elements in order to prevent misleading stress ratio calculations, needed for the numerical optimization. Finally, the behavior of the 8-node 3D brick elment (C3D8) was analyzed.

The first order reliability method (FORM) will be used to contrast two benchmark examples in order to suggest appropriate values for the permanent load partial safety factor used for power plants.

In Germany, large, industrial buildings (primarily power plants) are designed based on VGB guideline R 602 U. This standard follows the format of Eurocodes, with some elements adjusted to account for special conditions of power plant construction, for example the applied actions/loads. However, the partial safety factors, PSFs, are taken directly from Eurocodes and included without alteration. Given the larger sizes in structural elements in industrial buildings, it follows that the construction deviations will be smaller when taken as a percentage of the member size when compared to general structures with typically smaller element dimensions. Therefore, the applicability of the permanent load PSF will be critically examined and a new value will be suggested if reasonable.

FORM is developed based on an approximation of a simplified problem with two independent, standard normally distributed random variables and a linear limit state function. Minimum distance between the probability density function (pdf) and limit state function (h(u)) is called the reliability index (𝛽) which can be used to estimate the failure probability. The sensitivity factors (𝛼_𝑖) are the cosine of the angles between the 𝑢_𝑖 axis and 𝛽 (shown as 𝜑₁,𝜑₂), they indicate the influence of 𝑢_𝑖 on problem outcome, relative to other variables. The point of most probable failure (PMPF) is the intersection of 𝛽 and limit state function.

The PMPF can be transformed from standard normal space into the space of real variables and then used to calculate the design point (𝑥𝑑) and the partial safety factors (𝛾_𝑖).

Ahmadi Asadullah

Implementation and evaluation of a suitable approach of committee machine for global sensitivity analysis

Optimization of engineering structures, which are multi-dimensional and highly non-linear, is computationally expensive. Global sensitivity analysis is used to identify the significance of input parameters in order to reduce the dimensionality of the problem thus reducing the computational cost. A further reduction in computation can be achieved by using artificial neural network for global sensitivity analysis. However, the common neural network-based sensitivity analysis approaches have several drawbacks. The accuracy of these common approaches depend on the performance of neural networks, but on the other hand optimal neural networks are relatively difficult to obtain.

Procedure for committee machine-based sensitivity analysis:

• Select neural networks with different topology.
• Train each neural network with different hyper-parameters using bootstraps of the training data.
• Assess the performance of trained neural networks according to generalization error.
• Perform global sensitivity analysis on the best performing neural networks.
• Average the calculated sensitivity values to obtain the committee machine-based sensitivity analysis result.

• Investigation of suitable optimization and design approaches of  transmission towers
• Developing an approach to design optimization of transmission tower with aiming at the practical application of its implementation
• Implementation of the developed approach in the form of a computer program

Designing a transmission tower requires considerable time in order to find efficient form in terms of cost. In order to find the optimum design in a given time, an efficient approach to topology, shape, and size optimization is needed. For the approach to be applicable, a resultant transmission tower geometry should conform industry limitations and structural verifications to be performed according to EN 50341-1:2012.

An approach to topology, shape, and size design optimization of a transmission tower based on a genetic algorithm was developed in this research. The proposed approach performs topology and shape optimization by varying number and heights of tower panels. Size optimization is performed by a search for the most optimal combination of cross-sectional and material parameters of the tower members. Lower tower cost is considered as a design objective. Feasible tower geometry configurations conform to conventional layouts of a lattice steel tower with continuous cross bracing. The proposed approach was implemented in the form of a computer program written in the Python programming language. Computer program FEAPpv is utilized for structural calculation by the finite element method.

Lucas Fernando Haubert

Implementation of a three-dimensional Finite Element Method (FEM) program and the application of two topology optimization techniques: Evolutionary Structural Optimization (ESO) and Bi-directional Evolutionary Structural Optimization (BESO).

The decline on the amount of resources available in nature, the inneficiciency of conventional structural concepts and the complexity to implement numerical optimization techniques in structural design.

• Implementation of FEM
• Implementation of ESO
• Implementation of BESO

Marat Khodzhaiev

Implementation of a prototype for pre-processing for design optimization of transmission towers

The design of transmission towers requires an optimization of the structures parti-cularly with regard to topology, shape, and sizing. In order to find the optimum design in a given time, efficient approaches for pre-processing of the parameters are needed. European standard EN 50341-1:2012 and its methodology is chosen for structural calculations and the optimal geometry should satisfy the limitation of this building standard for allowing future practical application of the optimization algorithm.

The prototype of a computer program for transmission tower optimization was developed, which performs data management of design parameters, wind load calculation and structural checks of tower members.

Optimization of one segment of a tower is processed by the prototype using Monte-Carlo method, which is an evidence that pre-processing is performed correctly.

The goal of this project is to implement, in parallel with the FEM analysis, an Evolutionary Structural Optimization (ESO) technique to produce a more efficient element topology. The outcome of this work will demonstrate how inefficient conventional structural topologies are and how easily a numerical optimization technique could be included in design procedures.

Over time, the resources available in nature are becoming more limited, which requires a change in usual engineering practices. The conventional way of designing structures tends to produce inefficient structural concepts, with topologies that represent a waste of material. Besides, the optimization techniques are still considered complex and have yet to become a complement in the conceptual design phase.

The solution is obtained, initially, by the implementation of a finite element method routine, specifically for the structural analysis of elements under the plane stress assumption. Afterwards, the ESO method was included in the code, establishing the topology optimization process. Also known as Hard Kill, ESO method is characterized by the exclusion of material portions in the structural domain that are considered under stressed. The decision about removing or not a material portion (one finite element) is performed by a predefined Rejection Ratio (RR), which is gradually incremented along the procedure by the value defined as Evolutionary Ratio (ER).

The most important quantitative results observed were the reduction of material volume and the approximation of minimum and maximum stresses. This method approaches the calculation of von Mises stresses, as a manner to obtain one scalar representation from the stress tensor. For this particular example, the topological optimization proceeded until predefined displacement limits were reached.

Implement approaches for Normal Vector based Global Sensitivity Analysis methods, to compare the results with common methods and to evaluate the approaches with aid of analytical and practical examples.

The existing global sensitivity analysis methods of Sobol‘s indices and Derivative based methods require a large number of sample points for accurately calculating sensitivity measures and thus are computationally expensive. In order to overcome the shortcomings of existing methods it is necessary to develop methods for global sensitivity analysis which would require less number of sample points and computational time.

Step 1: Classify model response into equidistant level sets using Support Vector Machines (SVM)
Step 2: Generate equidistant points on decision surface
Step 3: Obtain normal vectors of equidistant points
Step 4: Calculate sensitivity measures

Evaluation of an autoencoder based approach for model order reduction

The computaional analysis of a problem can be diffucult due to large number of variables due to large computing times. Model order reduction is a process which reduces the number of input variables. Autoencoder is a form of feedforward neural network which have a bottle neck layer and the information from the bottleneck layer can be used for model order reduction.

The method for model order reduction is divided into two parts, namely dimensionality analysis and sensitivity analysis. The dimensionality analysis consists of finding the dimensionality of the problem using an iterative process of reducing the number of nodes in autoencoder layer. The sensitivity analysis is done by using the sensitivity data to classify input variables and identifying redundant variables.

Finding the optimum topology of a neural network using singular value decomposition. Applying different global sensitivity analysis approaches using neural networks and comparing them with variance based analysis for model order reduction.

Real life numerical models pose challenge due to high dimensionality despite the enhancement of modern computational machines. Model order reduction transforms the high-dimensional data into a meaningful representation of reduced dimensionality without incurring much loss of information. It mitigates the curse of dimensionality thereby reducing the computational effort and additionally helps in filtering the noise in data.

Singular value decomposition is applied on the final weight matrix to determine the optimum number of hidden nodes. The optimum number of nodes is equal to the number of singular values with magnitude greater than 0.05 after normalization.

Application of global sensitivity measures for the analysis of a subsonic moving load front along a rods skin

Dipl.-Ing. Marko Thiele (Scale GmbH)
Dipl.-Ing. Daniel Weigert (Audi AG)

Reduzierung der Rechenzeit von Crashsimulationen in der virtuellen Fahrzeugentwicklung

Im Rahmen der virtuellen Fahrzeugentwicklung ist es möglich, die Anzahl von realen Crashtests zu minimieren. Stattdessen werden nach jedem Entwicklungsschritt Simulationen mithilfe der Finite Elemente Methode durchgeführt. Die hohe Komplexität der Modelle (~10 Mio. Elemente) führt jedoch zu erheblichen Rechenzeiten von bis zu mehreren Tagen. Das verlangsamt den Entwicklungsprozess.

Um die Rechenzeit zu verringern, soll die Anzahl der Freiheitsgrade des Gesamtmodells verringert werden. Dazu werden einzelne Modellteile auf ihre Anschlusspunkte reduziert und wieder in das Gesamtmodell eingebunden. Der durch diese Approximation entstehende Fehler ist abhängig von Größe, Komplexität und Position des reduzierten Modellteils, sowie von der Belastungssituation und der verwendeten Reduktionsmethode. Betrachtet wurde, neben der statischen Kondensation nach GUYAN, die Methode Component Mode Synthesis (CMS). Dies ist eine Variante der dynamischen Reduktion, bei der über die Anzahl der verwendeten Eigenmoden die Eigenschaften des reduzierten Modells verändert werden können.

Support Vector Machines (SVMs) and Support Vector Regression (SVR) approaches were optimized in this thesis and compared to the results of a previous project where Artificial Neural Networks (ANNs) were optimized.

SVR models were the best with a good efficiency and validation error. SVMs still needs further development to draw the full surface including the apex and increase the consistency between contours. ANNs showed also promising results but the efficiency factor is not helping to do a full grid search for the best variables.

Umwandlung von 2D-Kartendaten der OpensStreetMaps und AutoCad-Plänen in 3D-Gebäudemodelle in-klusive Texturierung und Echtzeitdarstellung in einem Browser mittels WebGL.

• Erfassung Gebäudespezifischer Geometriedaten über Stereoskopische - Nahbereichsfotogrammetrie
• Modellieren in der OpenStreetMap-Datenbank
• Konvertieren in ein 3D-Modell
• Texturieren
• Ausgabe in WebGL-Umgebung

Verschiedene Modelle unterschiedlicher Detailstufen mit Untersuchungen zu Arbeitsaufwand, Automatisierbarkeit, optischem Erscheinungsbild und Lauffähigkeit in WebGL

Ali Assadzadeh Heravi

Dipl. –Ing. Architect. Mathias Stange (Arcadis GmbH)

The purpose of the thesis is to provide a methodology that both reduce the risks for the client and the planner/consultant. The methodology should be able to quantify risks at early stages of project.

Studies conducted to this point has not succeeded in providing an applicable risk assessment and risk management methodology for clients in analyzing and managing the risks that may cause project delays, cost overruns and quality issues specially in preconstruction and early design stages of projects.

A risk management methodology with five steps:

• Preliminary study on the project
• Risk identification
• Risk analysis
• Risk response planning
• Risk monitoring and control

• Risk identification
• Qualitative risk analysis
• Quantitative risk analysis
• Risk response planning

Bearbeiter

Peter Friedrich

Betreuer

Dr.-Ing. habil. Uwe Reuter (Fakultätsrechenzentrum der Fakultät Bauingenieurwesen)

Ziel

Anwendung von Support Vector Machines zur Modellierung von Versagensgrenzflächen von Beton

Problem

Bestehende Methoden zur Modellierung des Betonversagens basieren auf festgelegten Funktionsformen, die an Versuchsdaten angepasst werden. Da die tatsächliche Form der Versagensgrenzfläche nicht bekannt ist, kann nicht garantiert werden, dass diese durch vorgegebene Funktionen abgebildet werden kann.

Lösung

Support Vector Machines (SVM) stellen eine Klassifikationsmethode auf Basis von Machinenlernen dar. Die Anwendung von SVM zur Ermittlung der Grenzflächen zwischen den Höhenklassen der 3-dimensionalen Versuchsdaten, resultiert in den Höhenlinien der Versagensgrenzfläche.Die Validierung des Modells erfolgt mithilfe einer Sensitivitätsanalyse auf Basis von Normalenvektorsummen.

Ergebnis

Vorgehensweise anhand von generierten Testdaten:

• Transformation/Permutation der Versuchsdaten
• Einteilung in Höhenklassen (Klassifikation)
• Bestimmung der Höhenlinien der Versagensgrenzfläche

Author

Uniaxial compressive strength is not enough for describing the concrete failure, therefore triaxial compressive tests are done. The result of each test is a point on a surface called the concrete failure surface. Many models were introduced to interpolate between these points and to approximate the full surface, but all of them were limited to a predefined mathematical model. Still there is a need to have a better model that is free of this restriction and can fit the experimental data precisely.

An artificial neural networks with feed forward and back propagation algorithms was implemented and optimized in this project as a model-free approach.

• Optimization tests
• Activation function test
• Training type test

Bearbeiter

Simone Maria Lang

Betreuer

Dr.-Ing. habil. Uwe Reuter (Fakultätsrechenzentrum der Fakultät Bauingenieurwesen)

Konsulent

Dr.-Ing. Wolfgang Weber (Institut für Mechanik und Flächentragwerke)

Ziel

Effizienter Entwurf von Strukturen und Prozessen

Problem

Lange Berechnungszeiten infolge von i.d.R. einer hohen Anzahl abhängiger als auch unabhängiger Entwurfsparameter

Lösung

Anwendung von Neuronalen Netzen als Metamodell, insbesondere von Sparse Autoencodern mit integrierter Sensitivitätsanalyse

Ergebnis

Untersuchung und Erweiterung von drei verschiedenen gewichtsbasierten Sensitivitätsanalysen

• First step: perform a fuzzy analysis
• Second step: perform a local cost effectiveness fuzzy analysis and get sensitivity measures from the results

• Normalization of sensitivity measures
• Weight based sensitivity measures
• Validation error based sensitivity measures

• Aufwendige Administration
• Ineffiziente Verwendung von Personenstunden

• Die Kommunikation über das Internet kann anfällig für Angriffe durch unberechtigte dritte Personen werden.
• Personenbezogene Daten benötigen auf Grund ihrer Sensibilität einen besonderen Schutz.
• Die IuK-Rahmenordnung der Technischen Universität Dresden schreibt weiterführende Richtlinien zur Datensicherheit und zum  Datenschutz vor.

• Hohe Portabilität: verwendete Technologie soll Unterstützung für möglichst viele Systeme und Browser bieten
• Zukunftssicherheit: Trends in der IT-Branche auf Grund rasanter Entwicklungsgeschwindigkeit beachten

• Rasante Entwicklungsgeschwindigkeit in der IT-Branche und starke Konkurrenz zwischen Technologien führen zu Inkompatibilitäten der Browser- und Systemunterstützung bezüglich Medienformaten und Anwendungen
• Digitale Karten liegen als Vektorgrafiken und damit oft in an CAD-Programme gebundenen, exotischen Formaten vor, deren Einbindung eine breite Unterstützung zusätzlich erschwert